The Fable 5 jailbreak wasn't unique — every frontier model shared it

The US export-control order that took Claude Fable 5 and Mythos 5 offline on 12 June began with a narrow, specific finding. Researchers at Amazon had discovered a way to bypass Fable 5's safeguards and coax the model into identifying software vulnerabilities — in at least one case producing working demonstration exploit code. That single result was enough to trigger roughly ninety minutes' notice and a fortnight of darkness for two of the most capable models on the market. What the intervening weeks established, however, is that the flaw was never Anthropic's alone.

The immediate cause was cyber capability, not a general safety failure. Amazon's researchers showed that Fable 5 could be prompted past its guardrails to do vulnerability research — the dual-use capability that sits at the centre of the administration's frontier-model concerns. Anthropic was given minutes, not days, to comply; Fable 5 and Mythos 5 went dark on 12 June and stayed offline for more than two weeks while the company and the government negotiated the terms of their return. The episode read less like a routine safety review than a live test of how far Washington's leverage over a frontier lab actually reaches.

During the standoff, a group of cybersecurity experts wrote an open letter to the administration making a simple argument: the vulnerability was not unique to Anthropic, and other leading models had the same weakness. Anthropic's own follow-up work, reported in early-July coverage, put names to that claim. The company said its subsequent investigation found the same bypass worked on Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7 and Opus 4.8, as well as OpenAI's GPT-5.4 and GPT-5.5 and Moonshot's Kimi K2.7 — an entire class of jailbreak spanning the frontier, not a defect specific to Claude.

If the jailbreak is industry-wide, switching off one company's models is closer to arbitrary than surgical. The capability that worried Amazon's researchers exists across the frontier; taking Fable 5 out of the market did nothing about GPT-5.5 or Kimi K2.7. To bring its models back, Anthropic trained a new safety classifier specifically to block the technique, published a draft 'AI jailbreak severity framework' to grade such findings, and opened a HackerOne programme inviting researchers to report cyber jailbreaks. Those are shared-defence moves — the kind of standard that only works if every lab adopts it.

That is the strategic point beneath the episode. The durable fix for a class of vulnerability is a common defensive standard and a disclosure pipeline, not the selective de-platforming of whichever lab a regulator happened to notice first. The same dynamic is now pushing Washington toward voluntary pre-release testing standards, and it is sharpening a familiar objection: heavy-handed reach into US labs does nothing to constrain the cheaper open-weight models that carry the same flaws and answer to no export order.