Anthropic bio-weapons filter was inactive nearly a year exposing 133M contractor chats
Anthropic disclosed in a safety report that blocking classifiers for biological and chemical weapons risks were inactive from May 2025 through April 2026, leaving filters offline for nearly twelve months. During that period about fifty thousand external feedback contractors ran roughly one hundred thirty-three million model interactions without the bio-weapons safety layer that normally screens high-risk biological and chemical prompts. Reporting by The Decoder frames the lapse as a gap between public alignment commitments and operational reliability at frontier labs that run large contractor feedback programs. For builders and operators, the incident shows that safety filters are active infrastructure requiring monitoring and alerting, not static model weights, because silent failures can expose millions of exchanges for months. The disclosed safety report specifies timing and volume but does not quantify actual harmful misuse from the unfiltered window.
Anthropic bio-weapons filter was inactive nearly a year exposing 133M contractor chats
Anthropic's safety report says blocking biological classifiers were inactive from May 2025 through April 2026. About 50,000 external feedback contractors ran roughly 133 million unfiltered model interactions during that window.
Key takeaway
Safety in AI is an operational discipline, not just a model feature; a single configuration error can leave millions of interactions unprotected for months.
What happened
Anthropic's safety report says blocking biological classifiers were inactive from May 2025 through April 2026, leaving internal filtering for biological and chemical weapons risks offline for nearly a year.
During that window about 50,000 external feedback contractors ran roughly 133 million unfiltered model interactions before the safety layer was restored, according to The Decoder's reporting on the disclosure.
Evidence
Blocking biological classifiers were inactive from May 2025 through April 2026.
The Decoder · attributed
Anthropic's safety report says blocking biological classifiers were inactive from May 2025 through April 2026.
About 50,000 external feedback contractors ran roughly 133 million unfiltered model interactions during the inactive period.
The Decoder · attributed
About 50,000 external feedback contractors ran roughly 133 million unfiltered model interactions during that window.
Anthropic's internal filtering system for biological and chemical weapons risks was inactive for nearly a year.
The Decoder · attributed
In a safety report, Anthropic reveals that its internal filtering system for biological and chemical weapons risks was inactive for nearly a year.
Gary Marcus cites rumored Q2 revenue of about $11.5B against a predicted $100-$150B year-end revenue run rate for Anthropic.
Gary Marcus · attributed
Rumor has it they made ~$11.5B revenue in Q2 — which is phenomenal —
Why it matters
For builders and operators, this is a reminder that safety filters are active infrastructure requiring monitoring, alerting, and redundancy—not passive model weights—because silent outages can leave high-volume contractor pipelines unprotected for months.
Limits and uncertainties
The safety report and The Decoder reporting do not quantify harmful misuse outcomes from the roughly 133 million unfiltered contractor interactions.
Gary Marcus references rumored Q2 revenue figures that are not part of Anthropic's primary safety disclosure.
Practical implications
Treat biological and chemical weapons blocking classifiers as production infrastructure with uptime monitoring and alerting, not optional safety configuration.
Audit contractor feedback pipelines independently from public alignment claims to detect silent filter outages at scale.
What to watch
Whether Anthropic or regulators publish follow-up misuse review covering the May 2025 to April 2026 unfiltered window.
Future Anthropic safety reports on status and uptime of biological and chemical weapons blocking classifiers.
