Secret Claude model ‘better than all but the most skilled humans’ at hacking

Anthropic, the $380 billion AI giant responsible for the Claude tool, has a new AI model called Mythos that could become a crypto hacking nightmare.

Concerned about global panic if it were to release its frontier model too soon, Anthropic handed early “Mythos” access to JPMorgan Chase, Apple, Microsoft, and a few dozen other blue chip tech companies.

Unfortunately, Anthropic didn’t grant guest list access to any crypto company for its paternalistic Project Glasswing.

One Bitcoin developer asked Anthropic directly, “Why not cooperation with bitcoin/crypto projects?” Anthropic declined to reply.

JPMorgan Chase got a head start. Crypto didn’t.

A wing of glass preventing Mythos from hacking crypto

Anthropic’s Glasswing cybersecurity sprint is permitting 50-60 companies early access to its unreleased model that “can find software vulnerabilities better than all but the most skilled humans,” according to the company.

It is also donating $4 million worth of AI credits and is “committing” up to $100 million in AI credits for Glasswing.

According to Anthropic, which obviously has an incentive to praise the powers of its unreleased model for media and fundraising purposes, Mythos has strong reasoning and coding skills and is considerably more dangerous as a software hacking tool than most human developers.

Anthropic claims Mythos is “very autonomous” and has already found “thousands of high-severity vulnerabilities,” including bugs in “every major operating system and web browser.”

It withheld details about most of those bugs, except a 27-year-old bug in OpenBSD Unix software and a 16-year-old flaw in FFmpeg video software.

Curiously, Anthropic has published professionally staged and videotaped promotional materials in which Anthropic stakeholders sound alarms about Mythos’ capabilities.

For media purposes, it carefully selected the name “Glasswing,” which refers to the Greta oto butterfly whose transparent wings resemble glass.

Crypto, excluded from Glasswing, is particularly vulnerable

If Mythos’ threat is real, crypto software is particularly vulnerable to hackers with access to it.

Many implementations of crypto software are mostly or fully immutable, contain tremendous financial value, and have globally distributed deployment and upgrade cycles that prevent a quick defense.

Protos reported in December 2025 that even before Mythos, Anthropic had pitted its AI agents against 405 smart contracts.

Even with backdated knowledge and no internet access, its agents correctly predicted millions of dollars worth of available exploits on smart contracts which had gone live after researchers cut off the AI internet and knowledge access.

Anthropic’s AI agents also uncovered novel zero-day vulnerabilities in thousands of fresh contracts with no previously known flaws.

Those discoveries were before Mythos. According to Anthropic’s self-aggrandizing claims yesterday, Mythos can dramatically out-codes everything Claude has built previously.

Stifel analyst Adam Borg is convinced. “We read this as having the potential to become the ultimate hacking tool, and one that can elevate any ordinary hacker into a nation-state adversary,” Borg wrote about Mythos.

Anthropic says Glasswing partners will share their findings with the broader industry and patch major bugs prior to the public release of Mythos.