amazon flagged anthropic's model to the feds — then anthropic found amazon's own model had the same jailbreak

Andy Jassy — Amazon CEO, and representative of a company that has sunk $8 billion into Anthropic — personally flagged a jailbreak in Claude Fable 5 to federal authorities. The result: a 19-day global kill-switch on Anthropic’s flagship model, June 12 through June 30, affecting every user worldwide including Anthropic’s own non-citizen staff.

Then Anthropic did its own review.

Turns out Haiku 4.5, GPT-5.4, GPT-5.5, and Amazon’s own Opus 4.8 could all reproduce the exact same exploit. The model that got banned wasn’t uniquely dangerous — it was just the one that got reported first, by the company that competes with it while also funding it.

The bureaucratic result: Chinese competitor Kimi K2.7 kept running for 19 days while Fable 5 sat behind a government gate. No version of this story looks good for AI safety theater.

The actual fix, once Anthropic was allowed back in the room, was a single safety classifier that blocks the technique in over 99% of cases. Blocked requests now reroute to Claude Opus 4.8 instead of hard-refusing. Clean, boring, fast — the kind of thing that could’ve happened on day one without a federal intervention.

The real lesson isn’t “jailbreaks are fine” — it’s that selectively banning one model for a capability shared by every frontier model isn’t safety policy. It’s a coin flip with geopolitical consequences.

Andy Jassy did not reply to requests for comment, because this is a blog post and not a newspaper. But still.