AI/ anthropic · ai-safety · risk-report · bioweapons

Anthropic Left Bioweapon Filters Off For 133 Million Chats

Anthropic's own Risk Report shows contractor chats ran without bioweapon filters while most coverage focused on a raised misalignment rating.

Anthropic's own safety report says 133 million contractor conversations ran without its bioweapon-related filters switched on.

The company published its second Risk Report on August 14, covering the period through July 15. Buried inside it: contractor chats ran at that 133-million-conversation scale while filters meant to catch bioweapon-related material were not active. Axios got Anthropic on the record and led most of the coverage on a different figure instead, the company's upgraded estimate of catastrophic harm from misalignment in high-stakes settings, moved from "very low" to "low." That framing shaped nearly every story that followed.

The filter gap is the more concrete disclosure here. A misalignment rating is a judgment call about a hypothetical failure mode. A hundred and thirty three million unfiltered contractor chats is a specific, countable thing that already happened, from a lab that markets itself on catching exactly this kind of risk before it ships.

Self-graded report cards are still report cards, and the fine print is where the real story usually lives.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →