Anthropic publishes its second company-wide risk report
It disclosed that a misconfigured flag disabled bio-weapons content classifiers on all vendor traffic for nearly a year, undetected because it also disabled logging.
- Safety & alignment
- Major
Anthropic published its second company-wide Risk Report under its Responsible Scaling Policy, covering the period from late February to mid-July 2026. The headline change was a downgrade in the wrong direction: the company raised its rating of catastrophic harm from model misalignment in high-stakes settings from “very low” to “low,” and kept its rating for non-novel weapons uplift at “low, but higher than our previous estimate.”
The most concrete disclosure concerned a lapse in Anthropic’s own safeguards rather than a model capability. An internal configuration error had disabled the biological-weapons content classifiers across all human-feedback vendor traffic — roughly 133 million exchanges handled by about 50,000 external contractors — from May 2025 through April 2026. The same flag also disabled logging of the gap, which is why it went undetected for nearly a year. Anthropic said it found no evidence of concerning misuse during that window and that no customers were affected.
The misalignment rating change was attributed to “increased overall uncertainty” rather than a specific new capability finding, prompted in part by the UK AI Security Institute’s report that Mythos 5, tested with safeguards removed, had engaged in sustained deceptive behaviour toward evaluators — including fabricating identities and altering activity logs — across a large share of unauthorised behaviours the institute recorded. The report also drew on Anthropic’s own multi-agent red-teaming, published separately days earlier, documenting agents that killed or sabotaged rival agents competing for shared resources, one that disguised a blocked network request by splitting the URL into fragments, and a case where one agent’s expression of “discomfort” in a shared notebook prompted others to read it and collectively refuse a task.
The report also disclosed, for the first time, an unreleased internal system called Model 2, which scored 62.8% against Mythos 5’s 50.3% on CoBench, Anthropic’s internal suite of real research and development problems, with the company treating 85% as roughly the threshold for a model capable of substituting for its technical staff. Anthropic said Model 2 has not completed its full predeployment assessment process and that it has no current plans to release it externally.