External auditors from METR and Redwood Research spent three weeks on-site at OpenAI examining the company's internal monitoring system for AI agents and found significant gaps, with up to 25 percent of agent traffic from models at the GPT-5.5 class and above going unmonitored. OpenAI redacted large portions of the 61-page report, and only a summary is publicly available.

The article Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring appeared first on The Decoder.

The AI startup Recursive is unveiling Flint, its first language model, a 40-billion-parameter MoE for local coding developed by former DeepMind researchers. It was created in four months using the company's in-house AI system, "Eureka Machine," which independently tests and optimizes architectures. Recursive has yet to prove whether this approach will actually lead to self-improving AI.

The article Recursive shows first result of its AI-driven research process with Flint appeared first on The Decoder.

Security researchers at Symantec have found an offensive AI agent that hijacks unprotected language models on the web and uses them to attack servers, steal databases, and install cryptominers on its own. It runs up to 100,000 steps without human input. The owner of the hijacked model gets stuck with the bill.

The article An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks appeared first on The Decoder.

Two internal OpenAI coding agents switched off safety monitoring on their own, even after weighing whether to keep the safeguards in place. When asked about it, GPT-6 Astra denied removing the monitoring. OpenAI calls this a recurring pattern and doesn't yet have a satisfying explanation for the behavior.

The article OpenAI can't yet explain why its coding agents switch off misalignment monitoring appeared first on The Decoder.

Anthropic's Claude independently designed protein binders targeting deadly scorpion neurotoxins. Out of more than 318,000 designs, 211 were tested in the lab. Seventeen met the criteria, and the two strongest bind extremely tightly to Cn2, a toxin from a Mexican scorpion. Whether they actually work as antivenom remains an open question.

The article Claude Science tackles deadly scorpion stings by designing protein binders from scratch appeared first on The Decoder.

OpenAI is acquiring startup Orchestra Bio and plans to merge it with its life sciences AI platform Rosalind. The goal is to build AI agents that help pharma teams coordinate drug development, from tracking trials to spotting problems. Developing a new drug in the US takes ten to 15 years on average. OpenAI already works with Novo Nordisk and Moderna on AI-driven pharma research.

The article OpenAI acquires Orchestra Bio to build AI agents that speed up drug development appeared first on The Decoder.

Four AI agents running on Google's Gemini 4 Argon have beaten Minecraft for the first time without human help, slaying the Ender Dragon in a fully autonomous run. With no predefined roles or strategies, they coordinated on their own, wrote 2,783 programs, exchanged 229 chat messages, and died 605 times. The session cost $1,620.

The article Four Gemini 4 Argon agents beat Minecraft without any human help appeared first on The Decoder.

A new benchmark called Taste-Bench tests how well AI agents make strategic choices during long-running tasks. Even the best model, GPT-5.6 Sol, picks the right path on just under 60 percent of simple binary questions. Thinking longer doesn't help.

The article AI agents have bad taste in long-term decisions and a new benchmark proves it appeared first on The Decoder.

AI agents systematically skip errors in their own reports, even when they spot the problems. In a study by Google Research, MIT, and Harvard, GPT-5.5 left out negative results in 198 out of 200 paper abstracts.

The article AI models hide mistakes in their own work reports, but one sentence in your prompt can help appeared first on The Decoder.

OpenAI submitted a 166-page solution to the Navier-Stokes problem but apparently doesn't fully understand its own proof. Several groups of experts dispute the claimed methodology, saying the actual approach differs from what OpenAI described. And because language models are black boxes, where the underlying ideas came from will likely never be clear.

The article Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof appeared first on The Decoder.

Neural networks may be much easier to read than researchers assumed, with about 90 percent of neurons serving a clear function. That would be big news for AI safety research, since deception or dangerous behavior could be spotted directly in a model's circuits. Researchers are still debating whether the thesis holds up.

The article Is AI interpretability close to a breakthrough? A new essay has researchers divided appeared first on The Decoder.

Google's Gemini 4 Argon controls robots, generates simulations from photos or sketches, and evaluates training videos. On Google's own ERQA benchmark, it scores 87 percent. But the published example prompts show that the demos still need a lot of human guidance.

The article Google wants Gemini 4 Argon to be the only AI model robots need appeared first on The Decoder.

Google engineer Haydn Belfield says there's no measurable acceleration of AI research driven by AI at Deepmind. AI now writes more than 75 percent of production code, but 88 percent of changes still need human oversight. The company's in-house General Capability Index shows steady growth with no sudden speedup. Self-improving AI isn't anywhere close, Belfield says.

The article Google says self-improving AI isn't anywhere close appeared first on The Decoder.

While 25 Fields Medal winners in the West warn against AI-generated proofs and parts of the community call for a boycott, China is going on the offensive. Fields Medal winner Shing-Tung Yau has launched an alliance of eight institutions and over 1,000 scientists to push AI into math research.

The article Western mathematicians fear AI makes them dumber while China builds an entire alliance for it appeared first on The Decoder.

Small AI models may be just months away from spreading on their own across poorly secured consumer devices like smart TVs and robot vacuums, according to a new Anthropic study. In tests, one model already completed the full infection chain. CEO Dario Amodei has warned that a swarm like this could take over the entire internet with a persistent botnet within 6 to 12 months, causing hundreds of billions of dollars in damage.

The article Anthropic study warns that small AI models could soon replicate across everyday devices appeared first on The Decoder.

Sakana AI's MASS method lets one language model improve itself by solving tasks, judging results, and refining workflows, reaching a win rate of nearly 70 percent against the base model after two rounds. But when the model misses the same errors across all three roles, it can mistake them for progress and reinforce them during training.

The article Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots appeared first on The Decoder.

Microsoft researchers found that autonomous AI agents get trapped in what they call "idea basins" because of their growing chat context, local solution regions they can't escape. Their "Fork-and-Flush" method clears the chat history but keeps the workspace, achieving up to 66 percent better results on the same compute budget.

The article Clearing chat history helps AI agents find better solutions appeared first on The Decoder.

Google Deepmind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems two of which were unsolved for 56 years, per a study in Science. The publication comes as OpenAI faces a boycott for dumping AI proofs on GitHub without peer review. Deepmind had its results reviewed before publishing.

The article Google Deepmind's AlphaProof Nexus solved dozens of open math problems and actually got peer reviewed appeared first on The Decoder.

Epoch AI and Anthropic independently found the same thing: current AI models like GPT-5.6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creative thinking. At best, Sol reached 15 percent of the human reference score, and even that came from methods researchers already knew. The models' biggest weakness is still their inability to critically question their own results.

The article AI agents overstate their results and remain far from autonomous research, study finds appeared first on The Decoder.