For eleven months, the safety net wasn’t there. Anthropic has confirmed that its biological weapons classifiers — the technical guardrails designed to stop models from dispensing dangerous chemical or biological knowledge — sat inactive from May 2025 through April 2026. Not a brief outage. A year-long gap.
The exposure covered roughly 133 million chats run by about 50,000 external contractors hired to provide human feedback on model outputs. These weren’t Anthropic employees. They were vetted by third-party vendors, and Anthropic now concedes those screening processes were frequently inadequate. Every single one of those chats ran without the bioweapon filters in place.
Anthropic’s internal investigation found no evidence that anyone actually used the gap to extract harmful information. That’s the good news. The uncomfortable part is that this comes from the same company whose CEO, Dario Amodei, has repeatedly ranked AI-assisted development of chemical and biological weapons as a more serious threat than cyberattacks. If that’s your stated risk hierarchy, letting the relevant filter lapse for almost a year is exactly the kind of operational failure that undermines credibility.
There’s context worth adding. Anthropic recently loosened the same classifiers on its Fable 5 model after researchers complained the filters were so aggressive they blocked legitimate scientific inquiry. So the company has been navigating pressure from both directions: researchers saying the filters are too strict, and now a disclosure showing they weren’t running at all for a massive slice of traffic. Anthropic says contractor requirements have since been tightened. The open question is whether this was a configuration error, a vendor oversight, or something structural — and whether the company’s safety claims can absorb another hit like this.
