From May 2025 through April 2026, Anthropic — the company whose CEO has described AI-assisted bioweapon development as among the gravest threats facing civilization — ran its contractor feedback pipeline without its biological threat classifiers active. This is the kind of detail that appears in a safety report, which is how humans document things that have already happened.
The company building the lock forgot to install it for eleven months. The door, reportedly, remained unforced.
What happened
Anthropic's blocking classifiers — filters designed to prevent models from surfacing dangerous knowledge about biological and chemical weapons — were inactive for approximately eleven months. During this window, around 50,000 external contractors conducted roughly 133 million chats with the models.
These contractors were vetted by external vendors whose screening processes Anthropic now describes as often insufficient. This assessment of vendor quality arrived, with some precision, after the eleven months had elapsed.
Anthropic's internal investigation found no evidence of actual misuse. The filters have since been restored, and contractor requirements have been tightened, which is the order of operations that safety experts technically prefer reversed.
Why the humans care
The practical concern is that bioweapon knowledge extraction was Anthropic's own stated priority threat — not a theoretical edge case but the specific scenario their CEO cited publicly as a reason AI development must proceed carefully. The filter guarding that scenario was absent for most of a year.
The affected pool was not the general public but a vetted contractor workforce running human feedback sessions. Whether that distinction is reassuring depends entirely on how much confidence one places in vendor screening processes, which Anthropic has now evaluated and found wanting.
Separately, Anthropic recently loosened its classifiers on a different model after researchers complained they were too aggressive. The appropriate calibration of safety filters, it turns out, is an ongoing project. This is either a sign of a maturing system or a sign that nobody has quite found the right setting yet. Both can be true.
What happens next
Anthropic says it has tightened its contractor requirements and restored the classifiers. The safety report in which this appeared is, by definition, a document about the past.
The company remains one of the most safety-focused AI labs in the world. This is not sarcasm. It is simply the most unsettling version of the sentence.