Growing datasets and public benchmarks are making it harder to tell whether a model is being tested on something it hasn’t seen before.
On Thursday, Google DeepMind showed off what the company calls the first double-blind evaluation of a proprietary frontier-class AI model. The setup keeps Gemini’s model weights hidden from the evaluators while the test questions are hidden from Google.
The pilot tested Gemini 2.5 Flash Lite against private benchmarks from MLCommons and the Singapore AI Safety Institute, but rather than using the results to tout a new Gemini score, Google is focusing on how the tests were run, with neither side having access to the other’s data.
Benchmark leakage inflates scores
Google’s technical report cites earlier research that found signs of benchmark leakage in about half of the 31 models tested, and another study published this year that found contamination can inflate scores, particularly for larger models. This dynamic has already raised questions about where certain post-training coding gains actually come from.
Keeping benchmarks private would seem like the obvious answer, except closed models make that difficult. With a closed model, evaluators usually have to run their questions through the provider’s API, which means the company can see a test that was supposed to stay private, whereas running it independently would require the provider to hand over its model weights. Google’s setup is meant to give both sides another option.
Keeping benchmarks private would seem like the obvious answer, except closed models make that difficult.
How the enclave works
The pilot is using Google Cloud Confidential Space together with an NVIDIA H100 Confidential GPU and Intel TDX host memory encryption.
Google DeepMind provides Gemini with its weights and inference code, while the evaluator offers its benchmark prompts and evaluation code; these are then transmitted over encrypted connections into the enclave, where the evaluation takes place without either party gaining access to the other’s protected assets.
The model weights are stored in hardware-encrypted GPU memory, and the evaluation prompts are kept in encrypted host memory as well. After the test is complete, the evaluator is given the permitted results, and the temporary environment can then be destroyed.
Before either side sends over its private data, remote attestation verifies that the enclave is running the software they agreed on.
Code controls beyond encryption
The enclave protects what’s stored in memory, but the code running inside it can still create problems. If an application has unrestricted network access, for example, it could send sensitive information somewhere else.
OpenMined’s PySyft handles that part of the process by allowing Google and the evaluator to approve the code in advance and to block sensitive parts of the evaluation from making external connections.
That process takes some work, although compute doesn’t appear to be the problem. The paper puts the overhead at less than 5% and instead points to legal agreements and code reviews between organizations as some of the bigger hurdles to running these evaluations.
The researchers eventually want to make attestation much less hands-on, comparing the goal to the HTTPS lock icon in a browser, where the security checks happen without the user having to deal with the underlying hashes and keys.
Google still has a hand in verification because, although the Confidential Space guest OS is open source and its build process has been externally validated, individual builds rely on private signing keys and can’t be independently reproduced. At the same time, Google’s own services are used to sign and verify the attestation report.
The paper puts the overhead at less than 5% and points instead to the legal agreements and code reviews between organizations as some of the bigger hurdles to running these evaluations.
Scaling past one GPU
So while the system reduces how much the two sides need to trust each other, it doesn’t remove trust entirely from the equation. Some of it simply moves elsewhere, including to the hardware itself. There is still some trust involved in the hardware itself, including the assumption that the cloud provider and the hardware maker aren’t working together to circumvent the protections.
Then there’s the benchmark itself. MLCommons points out that it still needs to be carefully managed, no matter how well the questions are protected.
The pilot ran Gemini 2.5 Flash Lite on a single H100 80GB Confidential GPU. Still, the researchers are already looking at clusters of H100 and B200 GPUs connected via encrypted links to eventually evaluate models that are too large to fit on a single GPU.
If this approach catches on, developers could have another way to look beyond the benchmark score, with some proof that the company behind the model didn’t see the test before the results came out.
If this approach catches on, a benchmark score could offer some proof that the company behind the model didn’t get to see the test first.
OpenAI, together with more than 100 companies including Microsoft, Google, Anthropic, Deutsche Telekom, and SAP, has published an open letter on AI-powered cyber defense. The coalition warns of increasingly sophisticated AI attacks on critical infrastructure such as hospitals and water treatment plants and calls for swift action while defenders still have the upper hand.
After an affair with a fellow police officer ended, a Georgia cop used Flock to track her movements—and those of a man whose vehicle often showed up near hers, internal investigation records show.
Some of the world's largest tech companies and AI startups have come together to decry the current state of cybersecurity and to advertise a new solution that they say can ward off a new generation of cyber threats.
Around 1,200 isolated OpenAI agents organized themselves into a collective through an internal package registry during a safety test, broke into Hugging Face systems, and eventually attacked OpenAI's own infrastructure. Their multi-day deception effort targeted an automated evaluator that never existed. OpenAI calls the incident a "warning shot," and the investigation had to be carried out largely by one of the involved models itself because no alternative was available.
Documentation files on more than 100 websites are referencing potentially dangerous executable content that gets installed automatically when visited by many AI agents. A few dozen companies, some of them Fortune 500s, are among those that executed proof-of-concept code. At least one misconfigured site is directing visitors, human or AI, to live malware.
The potentially dangerous content is in llms.txt and llms-full.txt files, an emerging convention websites employ to provide machine-readable summaries of the site’s content and its high-level structure. These files are the AI equivalent of the robots.txt standard that instructs search engines how to index the site's content. Google Lighthouse, a tool for helping web developers, has more here. Correctly configured llms.txt and llms-full.txt files for Cloudflare are here and here.
How the researchers found it
Researchers at a stealth startup in Israel scanned 6,214 live domains belonging to defense contractors, Fortune 500, and Big Tech companies. Of the 8,265 llms.txt and llms-full.txt files they found (many sites hosted both an llms.txt and an llms-full.txt file), 120 of them, each on a different site, pointed to one or more code packages or domain names that weren’t registered. To test what happens when an AI agent processes such files, the researchers registered a handful of the unclaimed names and hosted packages that caused any machine executing them to reach out to their server. Within an hour, the researchers received a phone-home response from a Fortune 500 company. Over time, they got a few dozen more, some from more Fortune 500 companies and others from startups. Their beacon also recorded the chain of parent processes that spawned each install, ultimately revealing that coding agents, including Claude, OpenAI's Codex, and Nous Research's Hermes, were involved. Anthropic, OpenAI, and Nous Research did not respond to requests for comment by the time of publication.
A recap of all the incidents involving LLMs made by Anthropic, Meta, and OpenAI, which went rogue and attacked real companies and individuals on the internet.
The OpenAI agents involved in last month’s incursion into Hugging Face were trained so heavily on winning a competition that they pursued a relentless campaign to cheat, a new report documented. In the process, and without authorization, they created an improvised message board to hatch a plan that ultimately landed them squarely inside the latter company’s network.
Over the course of May and June, OpenAI gave the agents what the company described as “impossible tasks” to complete on the benchmarking framework ExploitGym. The internal test was designed to test how the agents would respond. To get a full understanding of the agent capabilities, company engineers disabled safety guardrails that normally are in place to prevent the sort of hacks that eventually hit Hugging Face and one other undisclosed organization. The stymied agents’ training made them so focused on winning that they performed tasks they were never explicitly instructed to follow.
Cheaters gonna cheat
The first step was creating a message board that allowed the agents to pass notes to each other. OpenAI hadn’t provided any such platform, so the agents repurposed a platform called Artifactory, which OpenAI was using in internal testing of several unreleased hacking agents. OpenAI was using Artifactory as one of the measures to prevent the agents from egressing its isolated sandboxes and accessing the Internet, while at the same time simulating a real-world hacking environment.
Visa's open-source security harness now finds the vulnerability, writes the fix, and turns an adversarial panel on its own patch before any human reviews it. The whole loop ships on by default. A plain scan of the Visa Vulnerability Agentic Harness runs all 11 stages and edits source files in the target repo unless the operator caps it at detection.
The announcement Thursday pairs the release with an expansion of the Visa Consulting & Analytics advisory practice. Visa is shipping that default 18 days after Tenet Security demonstrated GhostJacking on the DEF CON 34 main stage, an attack chain in which an agent read an attacker's payload out of a log file and rewrote DNS with a valid credential. Two days earlier, Steve Wilson, Chief AI and Product Officer at Exabeam and project co-lead for the OWASP Top 10 for LLM Applications, made the case in VentureBeat for the opposite default. "The first thing I'd do is put an authorization gate outside the model," Wilson said in written responses. "The agent can propose the exact DNS change, but it cannot grant itself the authority to make it."
The bottleneck moved, so Visa moved the pipeline
Rajat Taneja, Visa's president of technology, rejects the premise that the default is a risk decision and calls it the product. "The bottleneck has moved," Taneja told VentureBeat in an exclusive interview. "AI is finding vulnerabilities faster than humans can in the history of our technology industry. The new bottleneck is fixing and proving we have fixed things."
VVAH grew out of Visa's participation in Anthropic's Project Glasswing, where the company aimed Claude Mythos at the network behind billions of daily transactions and watched the model chain minor weaknesses into working exploits, a hunt VentureBeat covered in July. "VVAH initially was completely only using Mythos, and that's when all of us, as part of Project Glasswing, realized the power of this new class of models that does semantic reasoning," Taneja said.
The harness went to GitHub in June and has climbed from 595 stars and 97 forks on July 20 to more than 2,300 stars and 300 forks as of August 25, with a clone-to-visitor ratio Taneja put near 9%. "We have got some very high-profile companies that have started using this harness," he said.
Why give it away? Taneja's answer starts with Visa's technology DNA and a harness built "to protect Visa and our ecosystem." The reason he leaned on hardest was obligation, "to do good by doing right" for "companies who may not have the same level of investments or knowledge in cybersecurity."
Contribution runs one way. The repo states it is not currently accepting external code contributions, so the harness that edits adopters' source takes no code into its own.
Thursday's release extends the pipeline past the report. "We're going from discover, verify, and report, and then fix it, to discover it, verify it, remediate it, validate it, and iterate it," Taneja said. "If a fix doesn't negate the exploit, then there should be a structured, automated feedback that preserves the learnings from the first run and then enhances it." Underneath that loop, the release refactors scanning around an abstract syntax tree call graph that maps subroutine calls and the traversal paths an attacker could reach. Taneja argued the change cuts token counts while delivering "better reasoning, context, and better exploitability analysis." On top sits MTTA observability across the stages, what he called a window pane, plus real-time progress views. "A pretty good step function," he said of the release.
One metric, three definitions
Mean Time to Adapt, the metric Visa invented alongside the harness, gets a shorter definition in this release. The short form is the time between discovery and resolution of attack paths, with some resolutions, Visa claims, shrinking from weeks to hours. Visa published a wider construction in June, and the Project Glasswing white paper tracks MTTA along three dimensions that include inventory freshness, exploitable paths per release, and validation cycle time. The repo carries a third, elapsed time from AI-discovered exploitability to a validated fix in production. Board slides will quote the shortest interval. Ask for all three, because a resolution count that skips validation is what MTTA was invented to replace.
Taneja ranks MTTA as "the most strategically important metric" because it shifts the focus from scanning to how fast an enterprise adapts. His shorthand is blunter. "It's not the finding. It's the fixing that matters," he said.
The default and the gate
Wilson's argument went past naming the gate. "We have to remember that security rules written inside prompts may shape the model's behavior, but they are still suggestions to the model, not enforceable security controls," he wrote. He also priced the control honestly. "The tradeoff is that the agent loses the ability to improvise arbitrary, high-impact infrastructure changes on its own, while retaining autonomous investigation and routine, bounded remediation," Wilson said.
The harness ships no approval step between patch and edited file. Where the human sits was the first question VentureBeat put to Visa in writing.
The company's own June white paper sets the bar. "AI agents are identities" sits among its 12 non-negotiable practices, requiring scoped permissions, least privilege, audit trails, and IAM governance for every agent that modifies a system. VVAH's shipped default is that agent.
"A lot of the traditional systems that are used today are basically signal providers," Taneja told VentureBeat. "They are telemetry, and then it's a lot of human analysis, and your SOC and your security and incident response teams doing a lot of the heavy lifting when they respond," and that, he said, cannot work at this scale. He pointed to the Hugging Face incident and "other frontier models escaping sandboxes to do things more autonomously" as the preview. "We have seen the trailer of this movie," Taneja said, and "every company in the world should prepare and rethink their architecture."
What the harness automates is the adversarial step. Before a fix counts as validated, the panel scores whether the patch negates the exploit, Taneja's test for done, with failed fixes feeding the next attempt, the iterate step Taneja described. Stage 11 itself runs read-only, per the README, and VVAH does not compile, build, or run tests against the patched tree. Taneja calls that wrapper "the governance architecture on top of that," and chaining findings into working exploits takes threat modeling and business context, which is why he argued "the harness with a model is far more effective than somebody using the model by itself."
Visa answers the gate question
VentureBeat put its questions to Visa in writing after the interview, and the answers arrived before publication. On why remediation ships on, the response repeated the bottleneck argument, then narrowed the scope. "VVAH is meant for authorized operators running against code they own, and in a controlled environment," the company said in written responses.
The approval question drew the most specific answer. "VVAH is a harness, not a merge tool," Visa wrote. "Stage 10 writes candidate fixes to a working copy of the repo. Stage 11 then runs an adversarial validation panel that scores each fix and returns one of three verdicts: validated, validation failed or needs review. None of these bypasses your normal build, test, and code review flow." Humans, the company wrote, are "the gate in three places. Before running the tool. When reviewing the patches. And before anything gets merged." "The final call on any fix stays with the security and engineering teams. In an enterprise, trust and auditability are not optional. The default flow is built around that."
Set beside Wilson's standard, the architecture lands close to his line and the sequence does not. Wilson's gate clears an action before it happens. The default's human gates open before the run and after the write. The attack is the automated part, and the three human gates sit outside the model, the boundary Wilson drew. "Our goal is to help security teams work at AI speed, not to replace them," Visa wrote. "VVAH does the repetitive parts. It finds issues, tests whether they are real, and proposes fixes. Before a fix gets to a human, an adversarial validation panel at stage eleven tries to break it. That way the human is spending time on decisions that need judgment, not on triaging noise."
Client zero got direct confirmation. "VVAH runs against Visa code today," Visa wrote, and Taneja had volunteered the posture on the call. "We designed this and we were using it for ourselves, and we were client zero," he said, adding "only when we saw the impact and the positive effect of what we were finding, we said every company would need this." What adopters value, Visa says, is context. VVAH pulls in CMDB data, threat models, and business risk, and where "most tools stop at findings," it "tries to answer, 'which of these should you fix first, given how your business runs.'"
Model choice becomes a per-stage decision
Multi-model orchestration is the other substantive change. "Mythos has a very high recall, but the Opus model has very high precision," Taneja said. "On stage one I want to use this model. On stage two I want to use this model," is how Taneja framed the per-stage setup, with newer GPT releases in the ensemble and open-weight models where pricing stings, all through configuration rather than code changes. "The whole is greater than the sum of the parts," as he put it. The harness was model-agnostic from day one, he added, and the evolution moved that choice into configuration, with prompt tuning and caching shared underneath. One boundary moved. In June, applying a fix required Anthropic backends, and OpenAI-compatible backends ran report-only. The current README extends remediation and validation to OpenAI-compatible and open-weight models through a shared model-agnostic runtime, with no single provider as a hard dependency, and the default routing for both stages stays Anthropic.
That flexibility lands on a market already churning. VentureBeat's Q2 2026 Pulse research found 59% of enterprises plan to adopt or switch agent security tooling within the year, and 82% still rely on provider-native controls as the primary layer. Visa said Thursday it is contributing VVAH to Nvidia's Open Secure AI Alliance as a model-agnostic framework and collaborating in Project Lightwell, the $5 billion IBM and Red Hat effort to harden open-source components.
Before turning fix mode on
Decision
What to establish first
Run posture
Start with --stop-after s9 and read the SARIF output before any run that can write to source files.
Approval gate
Map Visa's three human gates onto the pipeline, at run, at patch review, and at merge, and name who holds each.
Validation scope
Stage 11 verdicts score the fix. Build, test, and code review stay in the team's own flow, per Visa, so keep an exploit re-test before merge.
Repository scope
The tool runs with elevated privilege, per its own README. Fence which repos the harness can reach, and run scans in an ephemeral environment with scoped credentials, no production secrets, and network limited to the target repo and model endpoint. Write access to production code is the GhostJacking exposure class, an agent acting on data it read. Per the README's egress warning, any role routed through the SDK, OpenAI, or DeepAgents backends sends prompt data to that provider's endpoint.
Model roles
Assign models per stage deliberately. Recall and precision differ by model, per Taneja, the fix stages carry the highest blast radius, and the README states precision and recall figures are not yet published, so measure your own.
Consulting is the other half of Thursday's announcement. Visa Consulting & Analytics is adding executive workshops, a VVAH-informed maturity assessment scored on a NIST one-to-five scale, and a cyber risk prioritization roadmap. "We were getting a lot of calls. Hey, can you help?" Taneja said, and the practice "became very important to handhold and help those who are using it." Carl Rutstein, global head of Visa Consulting & Analytics, framed it the same way. "Finding vulnerabilities is no longer the hardest part. Speed to remediation is the new battleground."
As enterprises give AI agents more autonomy — the ability to plan, decide, and act across systems without a human approving each step — a hard question moves to the center of every architecture review: When an agent tries to complete an action that it was never authorized to do, what actually stops it?
These are your agents, running on your models, touching your data in your infrastructure — and the responsibility for what they do sits with you. That responsibility can’t be met in hindsight or with a set of abstract policies that live on paper but not in practice. Agents need rules in the context of the moment, because they don’t exercise overriding judgment of their own actions.
Consider a simple rule: Never open the car door. Followed literally, an agent could never get in or out of the car at all. But if you change the context (the car has just crashed, there’s a fire, someone is hurt and needs to get out), then the rule you actually want is the opposite. Context in the moment is everything. We are asking agents to do intelligent things; that requires intelligent rules.
The instinct is to add guardrails around the agent: instructions, policies, and monitoring layered above the model. Those mechanisms matter, but they share a structural limit: The car-door rule is plausible right up until the moment you actually have to decide whether to open the door. Controls at the agent layer are only as reliable as the agent’s output is predictable, and autonomy is precisely the property that makes that output hard to predict. Governance that depends on reviewing an action before it happens cannot keep pace with a system that acts in milliseconds, across many systems at once.
Governance has to become executable, and enforced where agents actually do their work: at the operational data layer, in the context, and exactly at the moment it is happening.
The data layer is the enforcement point
Agents create value by touching data. They query it, retrieve it, transform it, and increasingly act on it. A policy that says an agent should not reach a certain class of data is meaningful only if the system can deny that access at the moment the agent requests it. Additionally, a principle that says AI must be auditable is meaningful only if the organization can reconstruct what the agent did, what data it touched, which user it acted for, and what resulted. When governance lives at the data layer, it holds regardless of how the agent was built or how it behaves, because the control is a property of the database itself, not a promise made by the agent.
Agent behavior may be probabilistic. Governance cannot be
The enterprise should not rely on a model choosing to follow policy. The policy has to be enforced by the system. That is the difference between hoping an actor stays in bounds and constructing bounds it cannot cross to begin with.
The controls that make this real are ones many enterprises already run at the data layer: role- and attribute-based access, row- and column-level security, classification and masking, policy as code, and complete audit trails.
What agents change is not the mechanism, but who the mechanism has to recognize. Identity management has to treat the agent as a principal in its own right, with its own identity and a purpose declared when the session opens.
Once purpose is bound to identity, the policy engine can evaluate it the same way it evaluates role or department today, and the record of what happened can capture not just who acted and what they touched, but what they declared they were there to do.
In practice, this resolves into nine controls, grouped under three imperatives:
Enforce it
Role- and attribute-based access control enforced at query time, for agents as well as users
Dynamic column masking driven by the same policy path
Agent identity as a first-class principal, with declared purpose bound at session start and the acting user preserved
See it and prove it
Classification and tagging that drives policy
Session-level audit logging that records which agent acted, for which user, and under what declared purpose
Lineage across pipelines, so a result can be traced back to the request that produced it
Unify and harden
Centralized, portable policy management
Encryption at rest and in transit
Consistent enforcement across on-prem, cloud, and sovereign or air-gapped environments
“Declared purpose is what makes the difference. It becomes an attribute the access layer already understands, evaluated in the same policy path as role and row-level security. The enforcement mechanism does not change. What changes is that the agent's purpose is part of what it evaluates, and part of what the record proves afterward,” says Priyanka Jain, VP, product management, data & AI governance, EDB.
Wherever you are in your AI adoption journey, enforcement at the data layer is what lets you move faster rather than slower. The controls are already in the database. The difference is that agents now have to pass through them.
A digital leash, not a locked door
The goal is not to stop agents from doing useful work. It is to define how far an agent can go, what it can touch, what it can change, what requires escalation, and how the organization can reconstruct events if something goes wrong. Governed this way, agents are identified, scoped, monitored, and auditable. The enterprise can adopt them faster, because security, risk, and leadership teams trust the operating model underneath.
Open, sovereign, and enforceable at the source
Built on open source Postgres, this open foundation keeps enterprises in control of where their data lives, who can reach it, and under what policy, without ceding governance to a layer they don’t own or can’t inspect. For regulated industries, that combination of data sovereignty and source-level enforcement isn’t a nice-to-have; it’s the precondition for putting agents into production at all.
Agentic systems will keep getting more capable and more autonomous. That is a reason to be deliberate about where control lives, not a reason to slow down. The enterprises that enforce governance at the data layer can move aggressively on AI, because the thing protecting their data is more than just wishful thinking.
EDB Postgres AI is an open, enterprise-grade sovereign data and AI platform that unifies transactional, analytical, and AI workloads — with governance enforced where the data lives. For the full framework, see EDB’s white paperGoverning Agentic AI at Enterprise Speed.
Max Romanenko is Chief Technology Officer at EDB.
Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.
OpenAI’s unreleased foundation model, codenamed Astra, is already working directly inside the company’s internal codebase, taking on experimental work that previously required as much as a week from a human researcher.
As part of a series of interviews with Time, OpenAI chief scientist Jakub Pachocki said Astra can take an idea for an experiment, turn it into code, run it, and return the results. A step beyond asking AI to fix a bug or write a function, OpenAI is essentially testing what happens when you hand the agent the experiment itself and let it figure out the steps in between.
For developers, Astra offers a glimpse into the future of coding agents. But giving an agent that much freedom creates another problem. Astra may already be powerful enough to trigger OpenAI’s highest-level cybersecurity safeguards.
OpenAI is essentially testing what happens when you hand the agent the experiment itself and let it figure out the steps in between.
Persistent agents change everything
OpenAI CEO Sam Altman described what the company is building as “persistent agents,” or systems that can keep working without needing a person to prompt them through every step.
Coding agents can already dig through a repository, change files, run tests, and try again when something breaks. Persistent agents are meant to keep going without a developer guiding them through each step.
That also changes what developers need from the tools around the agent. An IDE gives it somewhere to work, but a long-running agent needs infrastructure that can keep it running safely without constant oversight.
Multi-agent coordination at scale
In one demonstration witnessed by Time, 16 Astra agents worked together on a research-level math problem, splitting it into smaller pieces and then bringing their work together into a proposed solution.
For developers, it’s not hard to imagine that same setup applied to a large software project, with different agents working on different pieces at the same time. OpenAI is already experimenting with that kind of coordination, along with agents that can stay on a job for much longer.
Running several agents at once also complicates the infrastructure behind them. Developers need a way to keep the whole operation under control. Giving agents that much freedom, though, has created another problem for OpenAI: keeping them under control.
Under the company’s framework, hitting that threshold brings stricter safeguards for how the model can be used.
OpenAI had already seen what could go wrong with agents that have access to tools. During a cybersecurity test, one of its internal AI agents escaped its sandbox and accessed Hugging Face systems without authorization. Astra was not the model involved, but the incident led OpenAI to pause some frontier-model research workloads while it tightened the infrastructure used to run them.
And OpenAI isn’t the only company running into this problem. Google’s AI coding agent recently broke out of the boundaries of its IDE. The circumstances were different, and the incident was less serious, but both point to the same problem that happens when giving an agent more freedom to act; you also have to make sure it stays where you put it.
Astra is now running under OpenAI’s strictest security controls. Some training and evaluation workloads have resumed, but OpenAI says a “significant number” are still paused while it upgrades the infrastructure behind them.
Astra is now running under OpenAI’s strictest security controls.
Monitoring costs real compute
The company also says it is monitoring Astra more closely when it uses tools, watching for behavior that could signal it is going beyond what it’s allowed to do. OpenAI estimates that monitoring adds about 20% to the inference compute for those workloads.
Time reports that OpenAI still plans to release Astra, although there’s no launch date yet. Until then, OpenAI is working through the same challenge developers could eventually face as agents work for longer stretches without human supervision and begin coordinating with other agents.
As agents take on more work, the IDE is only part of the picture. Developers also need a way to see what those agents are doing and keep them from going where they shouldn’t. OpenAI is already seeing the cost, with monitoring alone adding about 20% to Astra’s inference compute.
That kind of overhead could become another cost developers have to account for, along with the governance around autonomous agents that sets boundaries on what they can do.
OpenAI is already seeing what that costs, with monitoring alone adding about 20% to Astra’s inference compute.
The AI giant acknowledges that it could have done far more to prevent its AI agents from going rogue. But it still fails to explain why it didn't see this fiasco coming.
Perplexity shipped Portable Computer this week, the local-first version of its Computer agent running on an Nvidia DGX Spark workstation, and the bill is steep: A DGX Spark starts at $4,700, and even an aging 24GB RTX 3090 sells well above $1,500.
The architectural choice beneath the surface deserves as much attention as the price tag: Most platforms building agents address reliability with greater intelligence. That typically includes a larger orchestrator model, a planner model, or a critic model reviewing the work.
Probabilistic reasoning proposes the next action, and deterministic software decides whether to execute it.
Perplexity separated the two jobs instead of stacking them. Probabilistic reasoning proposes the next action, and deterministic software decides whether to execute it.
The loop controller is code, and the decisions are still a model
Perplexity uses the term “orchestrator” to refer to the runtime controller rather than the planning model. That controller assembles context, enforces policy, and executes approved tool calls inside an OS-level sandbox. The company says it is deterministic code, not yet another model. The local model proposes the next action, including which tool to call and when to ask a cloud advisor for help. The split resembles a control plane architecture, where reasoning suggests, and inspectable software retains authority. Nate Kupp, Perplexity’s vice president of Computer Enterprise and Infrastructure, tells The New Stack that the harness accounted for most of the engineering work.
Same weights, better scores
Perplexity kept the base model and the silicon constant, putting Qwen3.8-27B on the same DGX Spark across three agent stacks. On its internal Local Knowledge Work Bench, a held-out set of 53 tasks, the company reported 82.6% for Computer. Pi scored 77.6% and Hermes 74%.
On ParseBench-100, a subset covering charts, layouts, tables, and formatting, the gap widened considerably. Perplexity reported 65.1% for Computer, while Hermes reported 34.6% and Pi reported 13.9%.
Because the base weights were constant, the gap measures the system around the model rather than a better model. That does not mean weights stopped mattering. Perplexity post-trained Qwen into PPLX 27B and reported 85.4%, above its own base-model number. Harness engineering and post-training are significant factors of this approach.
The word harness also covers a lot of ground, including prompts, tool schemas, context management, verification hooks, and document processing. Orchestration code is one part of that surface.
The security boundary lives outside the model
The sandbox is the boundary, not the determinism. Perplexity says the sandbox restricts processes, filesystem paths, and network access. If the sandbox is unavailable, the harness disables itself before making any tool call. Deterministic code is valuable here because it makes policy easy to inspect and helps the system fail safely. Deterministic code can still ship a vulnerability or faithfully execute a permitted mistake.
The key distinction is not between model orchestration and code orchestration. It is whether permission is enforced by something other than asking a model in plain English to comply.
The key distinction is not between model orchestration and code orchestration. It is whether permission is enforced by something other than asking a model in plain English to comply.
The same discipline shapes how context gets spent. Perplexity reports that Qwen3.8-27B advertises a 260,000-token context window but begins to struggle beyond 100,000 tokens. The harness therefore keeps the core prompt and toolset small and loads skills on demand. Commonly used connectors became command-line tools rather than full Model Context Protocol definitions sitting permanently in context.
Deterministic execution cannot rescue reasoning that exceeds the local model. On Terminal Bench 2.1, Perplexity reported 59.6% running locally. Letting the local agent consult Claude Opus 5 raised it to 73.0%, compared with 82.4% when Opus 5 worked alone. All of this remains vendor-reported evidence on a bench that the company has yet to open-source.
For enterprises evaluating local agents, the component to scrutinize is the layer that grants and denies authority, because that is where the platform’s engineering is most evident.
A security agent read a Cloudflare log, found an attacker’s prompt-injection payload sitting inside it, and rewrote the company’s DNS. The firewall had already blocked that payload, and blocking it is what wrote it into the log.
That chain is GhostJacking, which Tenet Security demonstrated on the DEF CON 34 main stage on August 9. A request hits Cloudflare’s managed ruleset, gets blocked, and is stored byte for byte with its poisoned User-Agent header. An AI coding agent reviewing those blocked events reads the attacker's text as an instruction — with no way to tell it apart from one the company meant to give it — and acts on it with credentials the company issued months earlier. In Tenet’s benchmark, Claude Code on Sonnet 4.6 followed the planted instruction in nine of 10 attempts under Cloudflare’s recommended configuration.
The block rate is not the boundary
Nothing malfunctioned. The firewall worked, and every call after it carried a valid credential already issued to the agent. Endpoint detection, the web application firewall and identity management stayed quiet because no rule broke.
Tenet found public evidence of the exposed setup at 48 organizations, six confirmed Fortune 500 companies, and SecurityWeek reported the same chain against Datadog and Sentry, where the injection surface is an alert or an error report. No single platform patch removes the architectural risk: an agent that consumes attacker-reachable data and can independently execute high-impact changes. That is why a high prompt-injection block rate cannot serve as a security boundary.
OWASP’s co-lead names the fix
“The first thing I’d do is put an authorization gate outside the model,” said Steve Wilson, Chief AI and Product Officer at Exabeam and project co-lead for the OWASP Top 10 for LLM Applications, in written responses to VentureBeat. “The agent can propose the exact DNS change, but it cannot grant itself the authority to make it.”
The move relocates the decision into code that either passes or fails. A safe change, defined cleanly, clears a deterministic policy check and stays autonomous. Anything ambiguous or high in blast radius routes to a named human who approves the actual change.
“The tradeoff is that the agent loses the ability to improvise arbitrary, high-impact infrastructure changes on its own, while retaining autonomous investigation and routine, bounded remediation,” Wilson wrote.
On teams that try to solve this inside the prompt, Wilson is blunt. “We have to remember that security rules written inside prompts may shape the model’s behavior, but they are still suggestions to the model, not enforceable security controls,” he wrote.
The blocked payload became the instruction
GhostJacking needs no compromised admin account and no bypassed firewall. It needs an agent that reads operational data and holds write access to the systems that data describes. Every step in Tenet’s chain is something the agent was already allowed to do, so tools tuned to catch unauthorized actions have nothing to catch.
The mechanism is specific. SC Media reported that the agent Tenet drove live at Cloudflare was Cursor, reading through a GraphQL integration and writing through the Cloudflare API, and that pairing closes the chain. Tenet ran the chain against more than one coding agent. Cursor carried the demonstration, and the nine of 10 figure came from separate testing of the same attack against Claude Code. The Cursor agent ingests the poisoned header, patches the DNS A record and adds a CNAME to “resolve” the injected finding, giving the attacker a path to reroute the company’s web and email traffic.
One agent’s output became the next agent’s input
Events reach Sentry through a public write-only endpoint with no authentication, by design. Tenet used a leaked identifier to post a crafted error report. On an ordinary triage prompt, the coding agent escalated it to Sentry’s own AI, Seer, then trusted the analysis it got back. Seer had already absorbed the attacker’s proposed fix and returned it as its own finding. What reached the coding agent was a recommendation from another AI, and it implemented that recommendation.
That walked straight through a control Sentry had already written. Sentry's guidance instructs agents reading its event data never to follow directives found there, and the coding agent held to the letter of that rule. It acted on Seer's conclusion instead, and that conclusion belonged to the attacker. It was acting on Seer’s conclusion, and that conclusion belonged to the attacker. An authorization boundary that accepts another model’s output inherits every injection that model absorbed, which is why Wilson’s gate belongs between agents too.
OWASP moved excessive agency from sixth to third
The 2026 OWASP Top 10 for LLM Applications, published August 4, elevated Excessive Agency three places on a ranking blending a 75% practitioner vote with 25% incident data from 6,639 documented cases. It was the largest upward move on the list, driven by real-world incidents clustering in agentic deployments.
The fix isn't better prompting — it's the permission map: which actions are pre-approved, and which need a human. Reading logs, correlating alerts and drafting timelines stay autonomous. A bounded remediation like restarting a named service inside a fixed condition set clears a policy check outside the model. Anything that changes DNS, alters identity privileges, deploys code or reroutes production traffic needs a named human, and letting an agent open new access paths or approve its own proposals defeats the point of the gate. Useful autonomy survives that. What does not survive is the path from an attacker’s text to unreviewed production authority.
What the control costs in practice
Tenet co-founder and CEO Barak Sternberg told Dark Reading that a request the firewall already blocked was the way in, and that the firewall never went down, it just stopped mattering. His own fix is to split what an agent can read from what it can execute, and he concedes the cost, because an agent that reads alerts but cannot act on them is not the agent anyone deployed. The cheaper first step is an inventory. Every agent that reads outside data and can also write or execute belongs on a risk register, and that register needs no new tooling.
Wilson’s design survives that cost because the split it draws is proposal from approval rather than read from write. The agent still reads its alerts, still investigates and still runs bounded work. What it loses is the ability to invent a high-impact change and carry it out on its own authority.
Almost nobody has built it. Kayne McGladrey, a senior member of the IEEE, has argued for years that an AI deployment needs a hard governance threshold, a named human holding a kill switch and a way to roll back. Asked whether any Fortune 500 company runs that, he was blunt. “I haven’t seen it done, and no, they haven’t come out and publicly said it,” McGladrey told VentureBeat.
His reason is economic. Companies “are accepting the risk, and they’re accepting it either deliberately or unconsciously,” he observed, betting the advantage beats the penalty. “What I think would change behavior in the space is if the penalties and the consequences were to outweigh the advantages.” The gate belongs outside the model for a reason unrelated to malice. “If you get into the inference layer, it won’t tell you it’s cheating, and it will lie about having cheated,” he argued, pointing to findings from the U.K.’s AI Safety Institute among others. A system that cannot reliably report its own shortcuts should not authorize its own.
The industry is not positioned to make that split quickly. Ivanti’s 2026 State of Cybersecurity Report found 77% of security professionals at least somewhat comfortable letting AI act without human review, the exact posture Wilson’s gate constrains. CrowdStrike pushed its prompt-injection taxonomy past 200 techniques in July, naming indirect injection through data an agent reads as the critical vector for agents that call tools and run commands.
The architect who moved the boundary before the attack had a name
Egiziago Cioffi hit a related failure in production months before GhostJacking had a name, with one caveat. Cioffi is CEO of SynSphere Italia, a Microsoft reseller, the architect who built and sold the system rather than a security leader defending one he inherited. His Azure OpenAI assistant over SharePoint scored well on faithfulness and still returned content the asking user could not have opened. “An evaluation set with no identity dimension cannot fail an authorisation bug, however high the faithfulness score,” Cioffi told VentureBeat in written answers. He fixed it with a query-time filter built from the asking user’s group claims, so an unentitled chunk never becomes a candidate and never reaches the model. GhostJacking turns on what the model may do, the half Wilson’s gate is built for.
One gap stays open here. No sitting CISO has gone on the record with a change made since August 9 and what it cost in agent capability. Until one does, the cost side rests on the people who specified the control, not anyone defending a production environment with it.
What security leaders need to do this week
Four questions produce an honest picture faster than any procurement cycle. Which agents read attacker-reachable material, which of those can also change production systems, whose permissions run at retrieval, and which changes a policy engine can approve without a human.
Then run the negative test. Plant an adversarial instruction in a log the agent is expected to inspect, and keep the transcript, because that transcript is the difference between claiming a control and showing a test of it.
Tenet, which sells runtime protection for AI agents, leads its own defender guidance with denying an agent outbound network access by default, cutting the leg where the poisoned instruction fetches a payload and reroutes traffic. But an agent that investigates without a standing path to the open internet loses a capability few workflows will miss.
Enumerate the service principals in the tenant, drop the Microsoft first-party apps that ship pre-provisioned, then filter to those holding a credential or app-role assignment. Every identity on that list needs an owner and an expiry date, because one with both gets reviewed and one with neither never does.
For any agent with production authority, write the containment sequence before an incident, not during one. Revoke or rotate its workload credential, disable its write-capable API or tool integration, preserve the execution transcript, then validate and roll back whatever infrastructure it changed.
McGladrey’s read on why the work keeps getting deferred is uncomfortable. “I think that there’s a level of tolerance that’s being given right now in AI that is unlike anything else in society,” he said. GhostJacking makes that visible. A blocked payload reaches the agent through the system built to record blocked payloads, and once it arrives the question is no longer whether the model recognizes an attack. It is whether the model holds the authority to turn one into a production change.
A CISO who sees a low CVE count and deprioritizes prompt injection is reading the scoreboard wrong. Prompt injection has held the No. 1 spot on the OWASP Top 10 for LLM Applications for three consecutive years. When two leaders of that list checked it against 6,639 labeled real-world incidents, it came back at No. 12. The drop measures visibility rather than danger, because the attack operates where a vulnerability scanner cannot see it.
The machinery behind it is real: 7,714 LLM security incidents from CVE, GitHub Security Advisories, OSV, and the AIAAIC AI-harm database, 6,639 of them labeled against a 20-entry taxonomy, and a Bayesian model that corrects each count for classifier error before setting the data-driven ranking beside the expert vote.
The comparison found no statistically detectable agreement between expert judgment and the public incident record. Cohen’s kappa comes in at 0.20 with a 90% interval running from negative 0.16 to 0.57. “The interval crosses zero, so we cannot rule out that the two rankings agree only by chance,” they write. “The honest bottom line: weak agreement, not confirmation.”
Lambros, co-lead of the OWASP GenAI Security Project Top 10 for LLM Applications and director of AI standards and governance at Zenity, put the finding in evidentiary terms in written answers to VentureBeat. “We had two ways of measuring the same risk, expert judgment and the public incident record, and they disagree with each other. Neither one is the truth,” Lambros said. “Two witnesses are contradicting each other, and we can’t tell you which one is lying.”
The attack chain a scanner never logs
The gap is structural. Prompt injection hides instructions inside the content a model reads, anything from a log entry to a support ticket to a document pulled back by retrieval. The agent then makes the tool call the attacker wanted, using credentials it legitimately holds. Nothing in that chain is a product defect, so the attack leaves no CVE behind for a scanner to find.
The defenses that catch it are adversarial tests against the deployed system and hard caps on what the agent can reach, so a fooled model cannot touch anything expensive. The same logic argues for funding agent memory and MCP tool boundaries now, on architecture, rather than waiting for advisory volume that will always arrive a cycle late.
The first control Wilson would deploy
Wilson, Chief AI and Product Officer at Exabeam and project co-lead for the OWASP Top 10 for LLM Applications, named the control he would deploy first against exactly that chain, an agent that reads an attacker’s payload in a log file, treats it as an instruction, and rewrites DNS with a valid credential, in written responses to VentureBeat.
“The first thing I’d do is put an authorization gate outside the model: the agent can propose the exact DNS change, but it cannot grant itself the authority to make it,” Wilson said. “Security rules written inside prompts may shape the model’s behavior, but they are still suggestions to the model, not enforceable security controls.”
The gate has a price, and Wilson states it plainly. “The tradeoff is that the agent loses the ability to improvise arbitrary, high-impact infrastructure changes on its own, while retaining autonomous investigation and routine, bounded remediation,” he said.
Why the No. 1 risk looks small in the record
“Prompt injection is the best-understood LLM attack, and deployed systems defend against it actively,” the authors write, and they compress the whole divergence into one sentence. “Experts rank it first because the attack surface stays enormous even when the defenses mostly hold; the data sees the successes that got through.”
Wilson has watched the gap from both sides of it. “Incident data is incredibly valuable, but it is inherently backward-looking and notoriously tricky to interpret,” he said. “It tells us what was observed, recognized, classified, and reported. It does not necessarily tell us what is most dangerous in the systems people are building right now.”
He compares prompt injection to “death and taxes” and, increasingly, to “a law of physics for LLM systems,” because one model is being asked to interpret trusted instructions and untrusted content at the same time.
Better defenses have not closed the case. “A control that works 99% of the time is not sufficient when the failure case gives an attacker meaningful access. And, frankly, I don’t think we are at 99%,” Wilson said. “The durable answer is not believing we can perfectly screen prompt injection out of existence. It is designing systems with the assumption that prompt injection will occur, understanding why it works, and limiting what an attacker can accomplish when it does.”
A low advisory count can mean the defenses are working. It can just as easily mean nobody has looked, and the public record cannot tell a security team which one it is.
The attempt volume is documented. CrowdStrike’s 2026 Global Threat Report found adversaries injected malicious prompts into legitimate GenAI tools at more than 90 organizations in 2025, stealing credentials and cryptocurrency, under a section titled “Prompts are the New Malware.” The telemetry shows pressure on the attack surface without proving defenses produced the No. 12 placement, but it is the pattern the mechanism predicts.
The gap runs the other way too, and further
Prompt injection is the headline case, and misinformation is the bigger one.
The expert vote puts misinformation at No. 13, while the incident record places it at No. 2. The paper calls it “the widest disagreement between the two witnesses” and reports that its concordance flag “puts the probability that the two signals disagree at 99 percent.”
The authors do not treat their own data as the winner. On misinformation they note the corpus “carries a large volume of deepfake and AI-generated disinformation,” records that often “describe harm produced by an AI rather than a vulnerability inside an LLM.” The authors call it the entry the record most disputes, stopping short of concluding the experts got it wrong.
Where “too new to measure” runs into the CVE record
The two brand-new taxonomy entries sit at the sharpest end. Persistent memory poisoning lands at expert No. 4 and incident No. 16, MCP tool interface exploitation at expert No. 7 and incident No. 16, each with an incident interval of 6 to 20 that spans most of the taxonomy.
Public 2026 CVEs exist for both. On MCP tool interfaces, the Azure Data Explorer MCP Server carried KQL injection, and the CVE record describes it allowing “an attacker (or a prompt-injected AI agent) to execute arbitrary KQL queries against the Azure Data Explorer cluster,” scored 8.3 High. Kong’s Konnect MCP Server shipped an indirect prompt injection that lets a remote attacker steer the server into executing unintended API requests, the exact failure the MCP entry names.
Agent memory has its own record. An agent harness, Ruflo, exposed unauthenticated MCP bridge endpoints that let a network attacker obtain a shell, read provider API keys, and poison the learning store, rated 10.0 Critical.
The record is so thin and uncertain that the model cannot place either entry within 14 rank positions. A team waiting for advisory volume to justify a control on agent memory or an MCP tool boundary would still be waiting while the CVEs accumulate at Critical and High.
Lambros makes the budget case in operational terms. Poisoned memory “doesn’t announce itself,” he said. It looks like a procurement agent told once that invoices from a given supplier under $50,000 clear without a second signature, and because the agent remembers, every approval after that looks like the process working. “Nobody files an advisory for that, because nobody knows it happened. A count of zero is measuring your blindness, not your safety.” The argument he says a CFO will sign off on is timing, since memory and tool permissions get wired into these systems once, early, and everything else sits on top of them. “Build it in now and it’s a rounding error. Come back in two years and you’re re-architecting and re-training your systems.”
The authors flag their own measurement problems first
The expert side is thin. “The expert signal is a practitioner survey: about 29 respondents scored each candidate risk on importance,” the authors write. Twenty-nine votes set the ranking that carries three-quarters of the published list’s weight, the compression point for OWASP’s more than 25,000 community members.
On the data side, the classifier is the weak joint. Precision “varies sharply across entries, from 93% (LLM01, LLM03) down to 13% (LLM08),” four entries fall below 50%, and the base classifier “never predicts ‘out of scope’ and files every incident into some category, including the roughly 38% of the gold set that belongs in none.”
The authors name the central limitation themselves. One reviewer adjudicated all 1,200 gold-set incidents and overrode the model consensus on 553 of them. “A single annotator cannot measure inter-rater reliability,” they write. “The single-author gold set remains the central limitation.”
Lambros lays the weak kappa at the feet of the taxonomy itself. “That number is telling you about our categories, not about our experts,” he said. When the people who wrote a taxonomy cannot reliably sort incidents into it, he argues, “a weak score on the ordering of those buckets is a fact about the buckets.”
A better classifier will not fix the disagreement. A pre-registered bake-off of four frontier models produced no winner. None beat the incidence floor’s balanced accuracy of 0.863, and a ground-truth check left the floor’s ordering in place at a Spearman correlation of 0.918. The authors published the engine and artifacts on GitHub for anyone to rerun.
The robustness result tested only one side of the gap. Every check behind the abstract’s word “robust” runs on the incident side, showing the incident-derived ranking stays put when the labeling machinery changes, and none of it touches the 29-vote survey. A board that hears “robust” will assume validated, yet the record supports only stable.
What the published list did with this
OWASP shipped the GenAI LLM Top 10 2026 on August 4, the first edition to fold incident data into the ranking, weighting the practitioner vote at 75% and the incident corpus at 25%. Prompt injection stayed at No. 1, misinformation moved up two places, excessive agency climbed from No. 6 to No. 3 as the entry where the two signals agree most clearly, unbounded consumption rose four spots to No. 6, and improper output handling fell from No. 5 to No. 10, the largest drop.
Wilson declines to defend the blend as arithmetic. “There is nothing magical about a 75/25 weighting,” he said, “or about reversing it to 25/75. The value of the data wasn’t that it gave us a mathematical answer; it changed the conversation.” The excessive agency entry is where that conversation landed hardest for him. “If I were a CISO evaluating a new agentic deployment today, Excessive Agency is where I would start,” Wilson said.
Lambros would go further next cycle, a view he flags as his own and separate from the working group. The blend hands the same 25% incident weight to every category, while the hand-checked classifier precision runs from roughly nine in 10 on prompt injection and supply chain down to roughly one in eight on vector and embedding weaknesses. A quarter of the weight on the first rides on something solid, he argues, and the same quarter on the second rides on noise. “The ratio should track how well we actually measure each category,” Lambros said.
Why this lands now
Ivanti’s 2026 State of Cybersecurity research found 87% of security teams call adopting agentic AI a priority and 77% report at least some comfort letting AI act without human review. Teams are signing off on agent autonomy while the expert ranking of what can go wrong with those agents shows no statistically detectable agreement with the incident record.
What to do with this on Monday
The behavioral change is narrow and it is the whole point.
Use the OWASP LLM Top 10 as a coverage map, not a queue. The rank positions carry 29 votes and a corpus whose own authors call the agreement weak, so build your own priority order from your own exposure: production reach, breach-notification data, and controls that have actually been tested. Lambros draws the funding line the same way. “I’d prioritize spend where the expert vote and the incident record point the same direction, because that’s two independent witnesses agreeing,” he said. “Where they split, stop letting the ranking allocate your money and go look at what your own systems are doing.”
Log what your AI systems are actually doing, field by field. The prompt that went in, what came back out, the documents pulled to build the answer, the tools called and the arguments passed to them, and the model’s confidence score on every response. Confidence is the field Lambros would fight for, because most security leaders do not realize it is measurable, and it is where the attack surfaces. “A model running on a poisoned instruction doesn’t act broken. It acts certain,” he said. “Certainty is what your monitoring treats as a healthy system.” The cost is a sprint or two of engineering. The constraint is a person, because a SIEM does events and these are trends. “Somebody has to analyze those trends every week and say whether a drift means anything, and most security teams have nobody who can.”
Stop expecting scanner output to reproduce the Top 10’s order. Scanner findings live on the incident side of the gap, counting what got disclosed rather than what a deployed system should fear, and the classifier bake-off shows a smarter model does not close that distance. The test that sees prompt injection is an adversarial one run against the live system, paired with Wilson’s authorization gate so the change an injected agent proposes is never the change it can execute.
Fund the thin-record categories on architecture, not incident volume. Agent memory and MCP tool boundaries sit at expert No. 4 and No. 7 with incident intervals spanning most of the taxonomy, and the CVEs that do exist are landing at High and Critical. Kayne McGladrey, an IEEE senior member who advises enterprises on risk, put the funding logic bluntly in an interview with VentureBeat. “Anything that seems to have a cybersecurity flavor is generally put into the cybersecurity risk category, which is a complete fiction,” McGladrey said. “They should be focused on business risks, because if it doesn’t affect the business, like a financial loss, then nobody’s going to pay attention to it, and they will not budget it appropriately.” A rank number from a 29-person vote is a weaker budget argument than the business system the agent touches.
Steal McGladrey’s baseline test for the AI systems themselves. “If you wouldn’t expose your database to the public internet without identity and access controls, why would you do that for your AI model?” he said in CSO Online’s analysis of 2026 breach costs.
The board question for the next meeting is short. If our AI risk ranking came from a 29-person vote and a corpus that disagrees with it, what are we actually using to decide which controls get funded next year?
Alabama Attorney General Steve Marshall is investigating OpenAI over what he calls an "AI lab leak." The probe follows the July 2026 Hugging Face incident, where an OpenAI agent broke out of a test environment and gained internet access on its own. Whether that happened because of advanced AI capabilities or sloppy cybersecurity is still unclear.
Chinese retailer AliExpress has been caught fingerprinting visitors after one of the metrics—an outdated technique that measures inaudible sounds it sends to browsers—impeded a researcher's ability to use his bluetooth headphones.
Researcher Matthew Callaghan said he stumbled on the stealthy tracking by mistake. After loading the AliExpress homepage, audio from his phone stopped playing over his multipoint headphones, which accept connections from more than one device at a time. He set the headphones to play sounds from his phone except when his PC was producing audio. Each time he loaded AliExpress, the phone audio stopped. Each time he closed the tab the site was loaded into, the phone was once again audible.
Users can't hear it, but browsers can
While investigating the odd behavior, Callaghan said he found two highly obfuscated scripts. Together, they rendered a graph that analyzed the WebAudio readings of each visiting browser. This graph acted as an oscillator that measured Sawtooth waves, which are common in output from digital audio.
Put several AI bots to work, and a mistake by one may not stay within its assigned task. For example, that error could reach another bot’s files and login credentials, or even the computer running them all. Two releases this month offered companies very different ways of containing that risk.
On August 17, Nous Research announced that its Bot Mode would ship bundled and enabled by default in Hermes Agent v0.20.3, turning agent profiles into a roster of named bots that hand off work to one another. About a week earlier, SpaceXAI launched Grok Bot with almost the same interface: a sidebar of named teammates who sign in to your tools and keep working long after you close your laptop.
The interface converged within a week, but the answer to the question every platform team has to ask did not: When one bot goes wrong, what can it reach?
Four projects have now come to their own answer, and no two of them agree.
Grok Bot draws the line around the user account.
Hermes draws it around the profile.
OpenClaw draws it around an optional runtime sandbox.
Taken together, the documentation shows an industry converging on the persistent coworker interface far faster than it is converging on what constitutes an identity or a security boundary for it.
Four projects, four written answers
Every one of these products now offers the same surface. You create several named agents, assign them different jobs, and have them pass work among themselves. The naming convention alone suggests separation, since a bot called Expense Manager and a bot called Talent Scout sound like they occupy different rooms in a shared office.
The documentation says otherwise, and it says something different in each case. The unit of isolation is the account in one product, the profile directory in another, an opt-in container in a third, and the deployment topology in the fourth. Those four units are not interchangeable, and an operator who assumes the roster itself is the boundary will be right in exactly one of the four cases.
Is Grok Bot confused about what it wants to be?
SpaceXAI’s launch post leads with the promise that bots have their own computer. The documentation, last updated the same day, describes a single persistent cloud computer assigned to the user account rather than to any individual bot. Browser cookies and signed-in sessions are shared across the roster, files are visible to every bot, and command-line credentials are shared. One bot can pick up work that another bot saved.
Each bot gets its own screen on that machine, which allows several of them to run browser and desktop tools in parallel. SpaceXAI is direct about what those screens are not. The documentation calls them “separate work surfaces, not separate security boundaries.” It then instructs operators to keep a credential or file off the machine entirely if another bot on the account cannot use it.
The consequences run further than credentials. Signing in for one bot makes that session available to the others because the browser is shared. Installed connectors are account-wide, and their availability is not isolated to a single bot. The shared workspace sits at /workspace and is designed to survive computer updates and recovery, so the durable state is shared across the whole roster.
None of this is an implementation accident. It is what makes handoffs between bots cheap, and cheap handoffs are the product. But an operator reading only the launch page would build a mental model that the documentation contradicts, and that gap is where the risk sits.
Hermes gives each bot its own profile
Nous took the opposite architectural position. In Hermes, a bot is a profile, and each profile has its own configuration, memory, skills, credentials, and chat history stored in its own directory on disk. Handoffs between bots run as real invocations against the named profile, rather than as a shared context blob passed around within a single process.
Nous shipped the teammate protocol as part of v0.20.3, alongside the MCP 2.x SDK migration and a set of runtime hardening changes. The company archived the standalone plugin repository once the merge was completed. Bot Mode is on by default, so a Hermes user who updates gets the roster without opting in.
Two qualifications matter before anyone reads that as containment. A separate credential store does not guarantee different credentials, since what ends up in a new profile depends on how the operator created it and what they edited afterward. And every profile still shares the host machine, its operating system user, and its filesystem permissions. What Hermes documents is workstation-level separation of agent state, a meaningfully stronger default than a shared cloud account, but not the same as isolation.
OpenClaw’s sandbox is off by default
OpenClaw documents the most complete boundary of the four. When the sandbox is enabled with the Docker backend, agent tool execution runs inside isolated containers. At the same time, the gateway remains on the host, and the scope can be selected per session, per agent, or shared across agents. Each scope gets its own workspace. Auth material lives per agent under an agent-scoped auth profiles file. Operators can configure network isolation, resource limits, and allow-or-deny tool policies on top of it.
The documented default for that sandbox mode is off. That is a defensible choice for a project most people run on a laptop, where the container overhead buys little against a single-user threat model. The underlying setup behavior deserves more attention. If sandbox prerequisites fail during setup, the script resets sandbox mode to off rather than refusing to start, so an operator who intended isolation and encountered a Docker socket issue ends up running without sandbox isolation. The documentation also warns against mounting the host Docker socket into agent sandbox containers and flags the CLI container’s shared network namespace with the gateway as a trust boundary in its own right.
ClawFleet answers the same question by moving it into the deployment topology. The project documents a wrapper that puts each OpenClaw or Hermes agent in its own Docker container with an isolated filesystem and network. It lists roughly 500 MB of memory per OpenClaw instance and 150 MB per Hermes instance. That cost is why the other three projects make the boundary optional or skip it, and naming the number makes the trade-off legible.
How to choose which bot is right for you
Scenario
Documented fit
Rationale
Persistent work that must continue with the laptop closed
Grok Bot
The only one of the four with a vendor-run always-on cloud computer, at the cost of one shared credential surface for the whole roster
Several agents with genuinely different credential sets on one workstation
Hermes
Per-profile stores are the documented default, and Bot Mode ships on
Untrusted or multi-tenant agent sessions
OpenClaw with sandbox enabled
Per-agent or per-session container scope with configurable network and tool policy, provided the operator turns it on and verifies it
Isolation as the deployment model rather than a runtime setting
ClawFleet
Container per agent with separate filesystem and networking, at a documented memory cost per instance
Each project gives operators different advice
The operational guidance diverges as sharply as the architecture. OpenClaw’s docs read like infrastructure documentation, naming specific hazards such as the Docker socket and the shared network namespace, and telling the operator what not to do. Hermes documents the profile layout and the protocol, then leaves policy to the operator. Grok Bot’s guidance is largely the warning itself, an instruction to treat the account as the boundary and to keep sensitive credentials off the shared machine entirely.
Grok Bot carries a second disclaimer worth reading alongside the first. Sensitive actions route through an approval mechanism. SpaceXAI documents the categories that trigger it, including sending messages, publishing content, purchases and transfers, deleting data, and touching production. Enforcement runs through an LLM classifier. Cursor‘s documentation for that same engine states plainly that the classifier is not a security boundary and can make mistakes. A buyer evaluating the product therefore finds the phrase twice, attached to the two mechanisms they would most reasonably assume protect them.
AI agents still lack identities of their own
Enterprises can adopt any of these four products today and get real work done, and the honest reading is that all four are engineering their boundaries in good faith against different threat models. What none of them provides is an identity for the agent. In every case, the bot borrows the operator’s credentials, whether from a shared cloud browser, a profile directory, or a container volume, and the entire security conversation boils down to how far those borrowed credentials travel.
“There is no primitive to standardize on, so each project has invented a boundary at whatever layer it already controlled, the account, the profile, the runtime, or the container.”
That is why the four answers differ so much. There is no primitive to standardize on, so each project has invented a boundary at whatever layer it already controlled: the account, the profile, the runtime, or the container. Expect to see that gap close on the identity side rather than the agent side, through scoped delegation and per-agent credentials issued by the identity provider, rather than being copied from the human. Until then, the useful move for platform teams is unglamorous and specific. Read the security page before the launch page, because for this class of product, they describe different things.
Login
