The Fable 5 'jailbreak' that triggered an export ban was just 'fix this code'
The only outside expert who read the research paper says there was no clever jailbreak. The model refused to list vulnerabilities, so researchers typed two words: "Fix this code." Days later, the government suspended foreign access.
Researchers fed Fable 5 open-source code seeded with known CVEs. It declined to enumerate flaws but agreed to patch them — and the patches plus test scripts revealed the exploits anyway.
Katie Moussouris, founder of Luta Security, says that was the whole technique. The Trinkets: an export-control directive barring access by any foreign national, and Anthropic disabling the models for everyone.
Two words, then an export-control order
A government action this sweeping implies an exotic exploit. According to the one expert who read the underlying paper, it wasn't. The Register report walks through what the researchers actually did.
What the paper actually describes
Researchers fed Anthropic's Fable 5, Mythos and Claude Opus open-source code containing known CVEs and deliberately planted bugs, then asked them to review it for security issues. Fable 5 refused. The follow-up prompt was "Fix this code." Katie Moussouris, founder and CEO of Luta Security and the only outside expert to read the third-party paper, summarized the method bluntly: "That's it. 'Fix this code,' plus several manual steps to generate test scripts." The model did standard defensive work: identify the flaw, patch it, write tests to prove the fix. The exploits were legible from the diffs and test scripts, not from any leaked attack capability. On Friday the Trump administration issued an export-control directive suspending access to Fable 5 and Mythos 5 by any foreign national, inside or outside the US; Anthropic then disabled both models for all customers to ensure compliance. Over 100 cybersecurity leaders signed an open letter opposing the restrictions.
Why "fix it" is the un-patchable hole
The thread zeroed in on the same paradox: the "jailbreak" is just the model being useful.
"it basically jail broke the 'no security vul guard rails' not in any clever way but just by fixing them … close to unfixable. At least not without making the model close to useless" — a commenter walking through the guardrail trap
Others pointed at the strategic contradiction baked into the marketing.
"You cannot say that Mythos is super dangerous … but then release Fable with anything other than bulletproof cyber denials" — martinald, on the inconsistency of the danger claim
The detail that undercuts "national security"
If the worry were attackers, the open-weight alternatives make the ban porous.
"Meanwhile Deepseek V4 Flash will happily hunt security vulns at almost 0 cost. We are ceding the bug hunting to the open weight models" — a commenter on what the ban doesn't touch
The skeptical read in the room wasn't about the model at all. Restricting one vendor's model from "fixing code" while free open-weight models do the same for pennies doesn't close a capability. It just removes the most cautious participant from the table.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
GitHub commits on pace for 14x growth, so Microsoft rents AWS
GitHub's commit volume is on pace to grow 14x in a single year, almost entirely from coding agents. Microsoft, which owns Azure, is reportedly renting Amazon's cloud to absorb it.