Anthropic's Safety Filter Was Broken for a Year

Anthropic recently admitted that their safety filters for biological and chemical threats were turned off for almost a year. This meant the AI models could answer dangerous questions without any safety guardrails in place.

During this time, contract workers processed roughly 133 million requests using these unprotected models. The company only discovered the error after the system had been inactive for many months.

This incident highlights the challenges companies face when testing powerful technology. Anthropic says the issue is now fixed and they are working to improve their internal safety checks.