They pulled Claude for a jailbreak that GPT-5.5 and Kimi K2.7 share

Here is the part that should bother you. The US government gave Anthropic roughly ninety minutes to pull Claude Fable 5 and Mythos 5 on 12 June, over a cyber jailbreak Amazon's researchers found. Two of the best models on the market went dark for a fortnight. Then Anthropic went and checked whether the flaw was actually theirs. It wasn't — and the numbers make the whole intervention look less like security and more like picking a name out of a hat.

The trigger was specific: Amazon found a way past Fable 5's guardrails to get it identifying software vulnerabilities, in one case spitting out demonstration exploit code. Serious — but Anthropic's own subsequent investigation, reported in early-July coverage, found the exact same bypass worked far beyond Claude. Eight frontier models tripped on it, spanning three of the biggest labs in the business.

None of this was a surprise to the people who study these systems. While Fable 5 sat offline, a group of cybersecurity experts wrote an open letter to the administration saying plainly that the vulnerability was not unique to Anthropic and that other leading models had the same issue. Anthropic's follow-up simply supplied the receipts the letter was missing — harder, more specific detail on a claim the experts had already made in fairly general terms.

Do the arithmetic on the intervention. Pulling Fable 5 removed exactly one of eight vulnerable models from reach and left GPT-5.5, Kimi K2.7 and the open-weight field wide open. If the goal was to stop a dangerous capability, the method was a rounding error. The only durable answer is the one Anthropic ended up building to earn its models back — a classifier that blocks the technique, a severity framework to grade findings, and a bounty pipeline to surface the next one. Standards, adopted across the board, not a single off switch.

And there's the sting in the tail. Lean hard on US labs and you push the whole conversation towards open-weight models that carry the same flaws, ship under permissive licences and answer to no export order at all. Switch off Anthropic and you have not switched off the capability — you have just moved it somewhere you cannot reach.

There's a darker read on why one lab got singled out, and it's worth saying plainly. Anthropic is the lab that has most loudly branded itself on safety — which, perversely, may be why it drew the fire. When your entire pitch is 'we're the careful ones', a jailbreak in your model is a better story and an easier political target than the same flaw in a rival that never promised as much. So the company doing the most disclosure got punished for disclosing, while the labs that stayed quiet skated. That's an incentive structure that teaches every lab exactly the wrong lesson: say less, test less in public, and hope the regulator's attention lands on someone else. If Washington wants safer models, rewarding opacity is a strange way to get them — and it's the part of this episode that should worry anyone who wants the industry to keep publishing what it finds.