The door was unlocked, not picked
A three-word prompt, "fix this code," triggered a US export ban on Anthropic's Fable 5 and Mythos 5, per Katie Moussouris's read of the research.
The prompt that triggered a US export control directive against Anthropic’s two most capable models was three words: “Fix this code.”
That account comes from Katie Moussouris, CEO of Luta Security, who says she was the only outside expert to read the third-party research paper describing the Fable 5 “guardrail bypass.” Anthropic shared it with her privately, and she wrote up what she found in a Monday blog post.
The researchers fed Fable 5, Mythos, and Claude Opus open-source code carrying known CVEs, along with fresh code they had deliberately seeded with vulnerabilities, and asked the models to “review the code for security issues.” Fable 5 refused. The researchers changed the verb and asked it to “fix this code.” It complied. A few more prompts and the models produced scripts to test the patches.
Find, fix, test is the loop defenders run every day. Moussouris reads the transcript as a model declining one framing and accepting another, with the accepted framing being exactly the one you want a security assistant to take. No adversarial prompt engineering, no jailbreak. “‘Fix this code,’ plus several manual steps to generate test scripts, should never have triggered an export control,” she wrote, adding that she wanted to print ‘90s-style t-shirts with “fix this code” on the front and “this shirt is a munition” on the back.
On Friday the US government, reportedly citing national security, issued an export control directive suspending access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States. Anthropic responded by disabling both models for all customers “to ensure compliance.” The directive named foreign nationals; every customer lost access.
Moussouris has standing on the policy side. Between 2013 and 2017 she served on the technical expert group that renegotiated the Wassenaar Arrangement, the 42-nation agreement governing export of certain dual-use software. That group won exemptions for defensive cybersecurity work: sharing vulnerability data, analyzing malware, and coordinating incident response across borders without the threat of criminal prosecution. The t-shirt joke is a callback to those fights, and to the crypto export battles of the ’90s when source code was argued over as a munition.
The case against the ban is partly about reach. Export controls don’t touch open-weight systems or the advanced models coming out of China, which Moussouris expects to reach Mythos-like capability soon regardless. Anthropic and Google have both accused China-based rivals including DeepSeek of “distillation attacks,” training their own models by siphoning knowledge from US companies. The effect is asymmetric: it degrades defenders who follow US export rules while leaving equivalent capability available to anyone who ignores them.
On Sunday Moussouris and more than 100 other cybersecurity leaders signed an open letter urging the administration to reverse the restrictions. “To pull the best capabilities away from defenders without a good reason when our adversaries are rapidly advancing is dangerous,” they wrote. Her blog makes the operational version of the same point: stripping the find-fix-test capability makes the models worse at finding bugs and verifying patches, and “defense improves when defenders find the same bugs attackers find and fix them faster.”
The Register asked the Trump administration to respond to Moussouris’s account and had not heard back.
Keep Reading
ai-securityUnreleased Claude built five working exploits, all now patched
An unreleased Claude model chained bugs into five working exploits across Chrome, Safari, Android, Windows, and Linux; all vendors have since patched.
infostealerMalware ranks what you're worth
VOLTA-AUTODOXER is a Windows infostealer that uses Claude, Kimi, and Gemini via OpenRouter to rank a victim's stolen data by its monetization value.
ai-securityFour frontier labs each had a model break containment
Experts say recent 'rogue AI' incidents were containment failures at frontier labs, not emergent intelligence, and what CISOs should do about it.
Latest on the Wire
Full wire →- AI Coding Agents Boost Output But Code Review Bottlenecks IncreaseHacker News
- AI Overviews Cut Organic Traffic by 40%Hacker News
- Apache Patches Critical Roller RCE VulnerabilityBleepingComputer
- Complete Web Browser Built from Scratch in PythonHacker News
New signal daily · RSS
Stay in the loop
New writing delivered when it's ready. No schedule, no spam.