Anthropic’s Bio-Weapons Filter Was Down for Nearly a Year #
Anthropic’s internal filtering system for biological and chemical weapons risks was inactive for nearly a year. During that period, around 50,000 external feedback contractors ran about 133 million unfiltered interactions with Anthropic’s models.
The disclosure covers large-scale external testing, but Anthropic’s report does not provide further details about the inactive system or the individual interactions.
Why it matters: Anthropic’s safety report puts the scope of the inactive control at around 133 million unfiltered model interactions conducted by 50,000 external feedback contractors.
Key Takeaways
- The inactive system was designed to filter biological and chemical weapons risks
- External feedback contractors ran the interactions during the nearly year-long period
- Anthropic described the affected model activity as unfiltered interactions