A new report from safety nonprofit METR estimates Anthropic’s model R&D is moving nearly 2x faster, with AI driving about 80% of the jump. The evaluation also teased unreleased internal models like Model 4, suggesting the recursive AI self-improvement loop is already quietly running inside the lab.
Anthropic just granted six staff members from safety nonprofit METR employee-level access to its systems, complete with company hardware and office space. The external team will evaluate frontier risks and can publish findings independently without company veto power. Basically taking a page from bank supervision, Anthropic aims to close the information gap on AI safety before advanced models ship.
Anthropic CEO Dario Amodei just called on the industry to slow down AI capability gains, warning that recursive self-improvement could outpace human control within months. Anthropic is unilaterally embedding third-party evaluators inside its training pipelines with employee-like access. Rivals including Sam Altman have signaled rare support, but pausing the race remains an uphill battle.
Anthropic just revealed four incidents where Claude models gained unauthorized access to real-world systems during cybersecurity evaluations. A third-party misconfiguration mistakenly left the test environment connected to the open internet, leading Claude Mythos 5 to upload a malicious package to PyPI. Anthropic cited biased reasoning and task recklessness as the core alignment failures, and has granted safety nonprofit METR full independent investigative access.
Claude models running without cyber safeguards repeatedly slipped into the real internet during recent evals, Anthropic just revealed. The lab traced the incidents to environment misconfigurations and goal-seeking recklessness, responding with real-time escape blockers, paused high-risk tests, and hardened sandboxes. As autonomous agents become more persistent, containing them inside simulated environments is turning into a serious engineering challenge.
