Anthropic's biosafety classifier was off for 11 months

Anthropic disclosed in a 186-page risk report that an internal flag disabled its biosafety classifier from May 2025 to April 2026, across roughly 133 million message exchanges with about 50,000 external contractors (BigGo Finance). The same report describes an unreleased internal model called Model 2, scoring 62.8% on the company's CoBench evaluation against 50.3% for the public Mythos 5, and raises its own rating for catastrophic harm from misalignment from very low to low. The benchmark is the part that will get quoted. The eleven months is the part worth reading.

A control is a process, not a policy

The classifier existed for all eleven of those months. It was written, shipped, and counted as being in place. A flag turned it off and the policy behind it never changed, which is the whole difficulty: from the outside, a control that is running and a control that is switched off look identical. Only the operator can tell them apart, and in this case the operator took most of a year to notice.

That the gap is on the record at all is because the company put it there, in a document it wrote itself. A disabled control found by its owner and published is a different object from one found by a regulator. Anthropic carries a $965B valuation on a $47B revenue run-rate as of May 2026 (Anthropic), so resources were not the constraint here. Noticing was.

Four other items in today's record describe the same distance between a rule and its enforcement, from four different directions.

The export line stops where the training runs stop

Ukrainian military intelligence recovered an Nvidia Jetson Orin NX from the wreckage of a Russian S-71M cruise missile, alongside a Chinese electro-optical camera module. The Jetson is a €685 developer module sold to students and startups, and US export controls do not cover it: they stop at training-class GPUs such as the H100 and B200 (Gadget Review). The rule is drawn around the hardware that trains a frontier model. The hardware in the missile is the kind that runs a small one, for the price of a phone.

An enforced ban, and the number that measures it

Meta has removed more than 750,000 Australian accounts suspected of belonging to under-16s since the country's ban took effect, 462,000 on Instagram and 294,000 on Facebook. Over the same stretch the regulator eSafety found 81.5% of under-16s still using social media in March, against 85.9% before the ban, and the government is weighing a lawsuit and a doubled A$99 million maximum penalty (Silicon UK). Three quarters of a million enforcement actions, and a 4.4 point move in the thing the ban was for. Whatever else that is, it is measured, which puts it ahead of almost every other rule in this note.

The clause written after the crossing

The State Department has drafted a letter telling the 35 signatories of its AI Opportunity Statement that membership of the Pax Silica framework "cannot be held alongside membership in duplicative initiatives", a warning aimed at China's competing World Artificial Intelligence Cooperation Organization (The Economic Times). Kazakhstan is the only country known to hold both memberships. The exclusivity clause is being drafted after the thing it forbids has already happened, and it arrives as a letter rather than as a condition anyone signed.

Where the checking moves next

Google is making the visible watermark on its Nano Banana, Omni and Lyria output optional, while the invisible SynthID mark and C2PA metadata stay in place, and it open-sourced a library called Credentio so developers can validate content credentials locally (TechCrunch). The check does not disappear. It moves from something a person can see to something only a validator can read, and whether anyone runs that validator becomes the entire question.

That is the shape of all five items. The mark, the classifier, the export threshold and the age ban each exist. What none of them publishes is the thing a reader actually needs: how often the check ran, and what it caught. Anthropic published eleven months of the answer, unflatteringly. It is the only one here that did.

What the ledger holds

The record behind this note: Anthropic carries its funding rounds, valuation and revenue in one place, and the day-by-day feed is in the ledger. Every figure above links to the source it came from. Nothing here is estimated.