Anthropic is restricting access to its most advanced AI model over concerns that the same capabilities making it exceptional at finding software vulnerabilities could enable unprecedented cyberattacks.
The company announced Tuesday that select Big Tech and cybersecurity firms will receive early access to Claude Mythos through Project Glasswing, a defensive initiative designed to bolster critical systems before potential adversaries exploit the technology.
Amazon, Apple, Microsoft, Google, Nvidia, Cisco, Broadcom, CrowdStrike, JPMorgan Chase, Palo Alto Networks and the Linux Foundation are among the initial partners gaining access to what Anthropic describes as "by far the most powerful AI model" it has ever developed. Approximately 40 additional organizations responsible for building or maintaining critical infrastructure will join them in using the preview version for defensive security work.
"Claude Mythos Preview demonstrates what is now possible for defenders at scale,"
Anthropic stated in a blog post first reported by Fortune. The company warned that adversaries will inevitably look to exploit the same capabilities.
On the CyberGym benchmark evaluating AI agents on vulnerability analysis tasks, Mythos scores 83.1%, significantly outperforming Opus 4.6's 66.6%. The model has already identified thousands of previously unknown software vulnerabilities at a rate far outpacing human researchers. Several discoveries had existed undetected for years, including a 27-year-old bug in OpenBSD, an operating system known for strong security.
Anthropic executives acknowledged extensive internal deliberation about releasing such powerful capabilities.
"We really do view this as a first step for giving a lot of cyber defenders a head start on a topic that will be increasingly important,"
Dianne Penn, Anthropic's head of research product management, told CNBC.
Project Glasswing participants will use Mythos to scan and secure their own systems while sharing findings with industry peers. The initiative includes $100 million in usage credits for participating companies and $4 million in direct donations to open-source security organizations.
The restricted rollout follows private warnings from Anthropic to government officials that Mythos makes large-scale cyberattacks considerably more likely this year. Company documents revealed in March described the unreleased model as representing a dangerous new level of AI performance in cyber tasks.
"Security research like this is especially important for open-source projects,"
Anthropic noted in its announcement. "By giving maintainers of critical codebases early access to these capabilities through Project Glasswing, we aim to help secure foundational internet infrastructure."
Anthropic plans eventual public release of Mythos-class models but maintains current restrictions to Project Glasswing participants only.
