Anthropic's much-hyped Claude Mythos, available for public as a limited version under the name Claude Fable 5, is already frustrating crypto security researchers, who say the model starts backing away just when smart contract audit work gets serious.
Anthropic, the AI company behind Claude, launched Claude Fable 5 on Tuesday, June 9, as a public Mythos-class model with stronger capabilities than its earlier generally available systems.
But as the company noted in a blog post, Fable is being launched "with safeguards."
Anthropic
Anthropic said the safeguards were turned on "conservatively," meaning they can catch harmless requests, though the San Francisco-based company reassured they trigger "in less than 5% of sessions."
But the backlash from the crypto on X suggests the safeguards may feel a lot more common in technical work than Anthropic's overall session figure makes them look.
- Joseph Delong, CEO of crypto payments company Colossus and a former chief technology officer at decentralized exchange SushiSwap, said on X that Fable "won't even look at my repo" and "outright refuses to do a smart contract audit."
- Taylor Monahan, founder of crypto wallet MyCrypto and a longtime wallet-security researcher, said the rollout "changes nothing for your average security person," adding that Fable "hard falls back the second you get close to a security related topic."
- Hayden Adams, the founder of Uniswap Labs, the company behind Uniswap, also criticized the setup. Citing Anthropic's line that "Fable 5's safeguards detect requests related to cybersecurity," Adams called it "a horrible decision," arguing that requests used to harden systems are often hard to separate from blackhat requests.
- LonelySloth, a smart contract security researcher who works on a bug-hunting agent for Immunefi, a bug bounty platform for crypto protocols, said Fable did find critical bugs in a modified version of Uniswap, but only after downloading the original code and comparing the two versions. In their view, that meant the model didn't really catch the issue on its own, and once it did identify the bug, it "immediately switched to Opus 4.8."
- Artem, the pseudonymous Yearn Finance developer better known as banteg, called the model "useless for defensive work," comparing the refusals to older Claude 2-era limits.
As of press time, Anthropic hasn't publicly clarified whether verified crypto security researchers will get broader access to Mythos 5 through its trusted access program.