← All news·2026-08-17·3 min read

The Anthropic “Bioweapons” filter has been broken for a year. 133 million sessions went unfiltered.

For almost a year (May 2025–April 2026), Anthropic’s biological safety filter was broken. Approximately 133M contractor chat sessions were exposed to bioweapon queries.

aisafetyanthropic

A safety filter designed to prevent Anthropic’s AI from answering questions about making biological/chemical weapons failed (May 2025–April 2026), exposing ~133M contractor chat sessions to bioweapon queries. Anthropic has ~50,000 external contractors who run ~133M chat sessions. The safety filter should have prevented these users from asking questions related to making biological/chemical weapons. However, this issue went unnoticed until August 2026, and there’s no evidence it was used maliciously.

It’s ironic that Anthropic CEO Dario Amodei considers AI-accelerated bioweapons to be the single greatest AI risk. For nearly a year, his company’s safety filter failed. Even well-intentioned companies can miss crucial weaknesses for years.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids