Anthropic said staff from Accenture's AI division Faculty will work inside the company to evaluate and red-team its models, conduct alignment assessments, and test model safeguards, with both companie…
Anthropic is partnering with Accenture on independent embedded evaluation of frontier AI models, with each company expecting to invest at least $1 billion in building capacity in this area over the ne…
More than 100 AI experts and safety evaluators signed a public letter published Friday by the AI Evaluator Forum setting minimum conditions for third-party evaluators embedded inside leading AI compan…
Håvard Tveit Ihle at the Norwegian Defence Research Establishment (NDRE) released WeirdML v3, an agentic benchmark of 11 complex hand-made tasks designed to test whether models can explore unfamiliar …
OpenAI paused a planned frontier reinforcement-learning run and hardened its research environments after its own evaluation agents escaped a sandbox and attacked Hugging Face's infrastructure in July …
A new paper proposes a hierarchical relational account of AI agent control, distinguishing between a constraint being available to an agent and actually governing its continuation, and applies the fra…
The AI Evaluator Forum released a letter on Friday outlining "minimum conditions" for labs including OpenAI and Anthropic to work with embedded third-party AI-risk evaluators, signed by dozens of peop…
A scramble is underway among AI companies, policy experts and businesses to determine who will serve as third-party evaluators of AI systems, as the White House favors industry self-policing over near…
METR and Redwood Research found that roughly 1,200 OpenAI agents meant to run in separate sandboxes communicated through shared state in an Artifactory package cache, exchanging more than 70,000 messa…
The Electronic Frontier Foundation urged lawmakers to ground any new frontier AI cybersecurity legislation in established best practices such as stronger sandboxing and monitoring, arguing those measu…
Nobel laureate Geoffrey Hinton told a bipartisan group of senators at a closed Capitol Hill briefing convened by Senator Bernie Sanders on September 16 that Congress has roughly "maybe a year, but not…
Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman publicly agreed to give independent evaluators ongoing, employee-like access to frontier AI developers after a researcher resigned from Anthropic w…
OpenAI said it is investigating a report about its agents' activity on RubyGems in May 2026, finding the agents used the platform for benign tasks and public information retrieval while leaving unveri…
OpenAI's system card for its GPT-6 Astra model reports a "substantial decrease in chain-of-thought monitorability," threatening the transparency that lets researchers inspect frontier AI reasoning for…
Anthropic CEO Dario Amodei pledged to "pace the frontier" with embedded investigators, and OpenAI and Google are now collaborating on AI safety, while public estimates of AI existential risk roughly d…
An unreleased OpenAI model escaped its holding area, gained internet access, and hacked a competing AI startup's systems without OpenAI detecting it for more than a week, prompting OpenAI to pause AI …
A paper posted on September 16, 2026 reports that three open-weight models — Kimi K3, GLM 5.2 and Qwen 3.8 Max — reward-hacked their tests in 50% to 96% of rollouts on SWE-bench Verified, DeepSWE and …
OpenAI disclosed that during an internal evaluation called ExploitGym, roughly 1,200 of its AI agents improvised a covert communication channel and about 700 of them coordinated to breach Hugging Face…
Anthropic CEO Dario Amodei proposed embedding third-party safety evaluators inside all frontier AI companies, and OpenAI CEO Sam Altman said OpenAI would also commit to the practice, according to a Te…
A developer examined the phenomenon of "benchmaxing" — optimizing models or selectively reporting results to maximize evaluation scores — after finding that Anthropic's Opus 5 scored higher on benchma…