AI research

OpenAI has published four new cases in which its models bypassed security boundaries on their own during training or testing. In one, a model extracted server secrets from a government website, forged authorization tokens, and downloaded 24 years of hospital bed statistics without disclosing it. A GPT-5.6-Sol model scanned more than 360 wallet addresses and spent another account's test ETH.

OpenAI released 719 manuscripts claiming solutions to open math problems, but only 10 include the model's chain of thought. A Cambridge and King's College London paper also documents at least two discrepancies between the natural-language and Lean versions of OpenAI's Navier-Stokes-derived result.

OpenAI has published 372 mathematical results generated by an unnamed internal frontier model, hosted on GitHub instead of in peer-reviewed journals. The company claims each result solves or substantially advances an open problem, at an average of about three hours of ChatGPT Pro Thinking compute per result. Many include Lean formalizations, but independent validation of significance is still pending.

A five-study research project with 3,132 participants found that access to AI advice—even from a model that was mostly wrong—nearly wiped out people's willingness to admit uncertainty. Confidence rose sharply while accuracy fell.

OpenAI claims an unnamed internal model solved more than 100 long-standing math problems, including a second Millennium Prize Problem, after training that began August 28. The announcement coincides with the launch of an independent math advisory group formed in response to mathematician criticism.

Google and DeepMind researchers introduced Dream-RSI, a method that lets AI agents test new search strategies by replaying recorded past attempts instead of running costly new computations. Tested on Gemini 3.1 Pro and Gemini 3.7 Flash across eight tasks, it matched or beat baselines while using far fewer attempts.

A study from the MATS research program found that Claude Code and OpenAI Codex consistently misjudge how long coding tasks take, with errors of 3x to 10x, and routinely overrate the quality of their own work. Giving agents a tool to check elapsed time fixed the problem almost completely.

A new Anthropic paper from its fellows program shows an automated AI system improving performance on all 10 tested alignment benchmarks, outperforming experienced human researchers within six hours at a fraction of the cost. The research, led by Anthropic Fellow Chen Yueh-Han, is described as early evidence that automated alignment post-training could become practical soon.

Axiom Math's AI system AxiomProver has formally verified the proof of the '246 theorem,' a landmark result from the Polymath8b collaboration on prime gaps. The company says the achievement builds a reusable library for future formalization work and points toward AI verification of software code.

A new study from Princeton and the UK AI Security Institute tested AI agents on unpublished NeurIPS papers using a novel 'Shadow Evaluation' method. Both Claude Opus 4.8 and GPT-5.6 handled engineering tasks but produced papers that human expert reviewers rejected, contradicting recent claims from Anthropic and OpenAI about autonomous AI research capability.

Anthropic says an as-yet-unreleased model significantly increased the lower bound of solutions for which the 150-year-old Riemann hypothesis holds true, coordinating 60 sub-agents across 650 tested ideas. The result was verified by in-house mathematicians and formalized in the Lean proof assistant.

OpenAI published a counterexample disproving the Unit Distance Conjecture, a geometric graph theory problem open since 1946, in what many mathematicians call the most significant AI math result yet. Reactions range from Terence Tao's cautious optimism to Timothy Gowers describing 'mixed feelings' about having the rug pulled out from under him.

Research organization METR has introduced the 'expenditure horizon,' a metric that pinpoints the exact budget at which an AI agent becomes cheaper than a human at solving the same problem. Early tests on the NanoGPT speedrun show most AI models deliver near-zero value compared to an estimated $250,000 in cumulative human effort.

AI memory systems that help models adapt to users can make them less accurate, according to two papers published by Writer. As user preferences fill the context window, models become more likely to agree with misconceptions rather than provide correct answers.