Anthropic is opening access to its most powerful artificial intelligence model Claude Mythos to roughly 150 new organizations, one day after the company confidentially filed for a public offering.
Key Points:
- Anthropic added roughly 150 new organizations across more than 15 countries to Project Glasswing, its access program for Claude Mythos.
- The expansion lands one day after the company confidentially filed for an initial public offering.
- Researchers warn the model could make cyberattacks easier even as Anthropic pitches it as a defensive tool.
Glasswing Expansion Spans 15 Countries
The company announced Tuesday that it is widening Project Glasswing, a controlled program built to find and patch software flaws before the model reaches the public.
The new partners sit in more than 15 countries. Each one must clear security requirements before gaining entry.
An initial group of about 50 organizations joined in April, among them Google, Microsoft, Palo Alto Networks, and JPMorganChase. Those early users have since flagged more than 10,000 high or critical software vulnerabilities. The fresh wave brings the total to roughly 200.
The latest cohort fills key gaps. It spans power, water, healthcare, communications, and hardware, sectors that lagged in the first round.
A single breach at one of these firms, the company estimates, could harm well over 100 million people.
Also Read: Study Finds Four Major AI Labs Use Incompatible Prompt Injection Metrics
Why Claude Mythos Worries Researchers
Researchers and government agencies have questioned whether models this capable could make complex cyberattacks easier to pull off. Anthropic argues the opposite, saying defenders need sharper tools before attackers gain them. It frames the rollout as a lasting edge for defense.
Independent testing has only fueled the debate. The U.K. AI Security Institute documented a 32-step simulated network attack the model ran on its own. Mozilla reported that Mythos found 271 vulnerabilities it then fixed in the Firefox browser, while startup Calif described an exploit chain aimed at Apple's M5 chips.
The company expects rival developers to ship comparable models within six to 12 months, possibly with weaker safeguards.
That gap, it warns, could bring more frequent attacks. In the meantime, Anthropic has released Claude Security, a tool on its public model that has patched more than 2,100 flaws in three weeks.
Claude Mythos surfaced publicly back in March. Draft company materials had leaked online, and Anthropic has since refused to hand the full model to everyone without firmer safeguards. On the prediction market Myriad, odds of a public release by the end of June have slipped to about 26.5%, down from 58.5% a week earlier.
Read Next: Durov Renames Toncoin To Gram, Coin Pops 10% On Roots Return
