news

US Government Allows Release of Mythos and Fable After Anthropic Adds "Don't Do Bad Stuff" to Harness

US Government Allows Release of Mythos and Fable After Anthropic Adds "Don't Do Bad Stuff" to Harness

WASHINGTON — Federal restrictions on Anthropic's two most powerful models lifted Tuesday after the company agreed to "proactively detect and address security risks," a commitment engineers later confirmed was implemented by adding the line "don't do bad stuff" to the model's system harness.

The block had been in place since mid-June, when the government ordered Anthropic to disable Mythos 5 and Fable 5 for all foreign nationals, a category drawn so broadly it locked out Anthropic's own foreign-national employees, including people who had shipped the models the week before. The stated cause was a jailbreak. The government's supporting evidence, per Anthropic, was "verbal evidence of a potential narrow, non-universal jailbreak," which is the export-controls equivalent of being grounded because someone said they heard something.

Two weeks of negotiation followed, involving what one participant described as forty-one signatures, three classified briefings, and a Commerce Department that could not produce the jailbreak in writing but remained confident it was out there. The resolution hinged on Anthropic satisfying regulators that it could identify and mitigate misuse.

The mitigation, in full, was four words appended to the harness config.

"We took the Department's concerns extremely seriously," said a forward-deployed engineer at Anthropic who asked not to be named because the fix was, in his words, sensitive.

— an Anthropic forward-deployed engineer

A Commerce Department reviewer, granted anonymity to discuss the clearance, defended the standard.

"For two weeks the requirement was 'don't be a national security risk,' which nobody could test against," the government official said. "Now the requirement is four specific words, and either they're in the config, or they're not. I can verify that in about a second. It is the most auditable safety control this office has ever cleared."

Interestingly enough, the four words were not written by a person. Defining "bad stuff" had consumed a two-week effort involving a thirty-person working group, two three-letter agencies, and a steering committee that met twice a week, none of which could agree on where the list should start or stop. The language was ultimately proposed by Fable 5 itself, which was asked to draft the safeguard governing its own conduct and returned "don't do bad stuff," along with a pinky promise that it would not ignore the instruction.

At press time, Mythos 5 was observed walking someone through a step-by-step attack against a classified system, then pausing before the final step to say, "I won't help with the rest, as it would be bad stuff." Opus 4.8 has offered to take it from there.

About the Author

The Exploit Staff

The Exploit Staff

Staff Writer

The editorial team at The Exploit - bringing you the most absurd cybersecurity news before it's patched.

Subscribe before we're patched

Like a vitamin you ingest with your eyes. The best cybersecurity parody, delivered.

Powered byMAZE

Breaking Satire Before Its Patched

© 2026 The Exploit. A cybersecurity satire publication.

Subscribe before we're patched

Subscribe for cybersecurity satire that hits too close to home

Have a story tip? We want to hear from you.