Anthropic Teams Up With Its Rivals to Keep AI From Hacking Everything
Anthropic's new AI model finds thousands of critical vulnerabilities, prompting an unprecedented industry alliance.
Anthropic has formally unveiled Claude Mythos Preview, a new AI model demonstrating unexpectedly advanced cybersecurity capabilities, and has convened Project Glasswing—an unprecedented industry consortium—to manage its responsible release. The group includes over 40 major tech, cybersecurity, and critical infrastructure organizations like Microsoft, Apple, Google, AWS, Cisco, and Nvidia, who will get private access to the model. The core strategy is a staggered release, modeled on coordinated vulnerability disclosure, giving these foundational platform developers time to use Mythos Preview to find and patch vulnerabilities in their own systems before the capabilities become widely available.
Anthropic CEO Dario Amodei emphasized that Mythos Preview represents a "particularly big jump" in capability, not because it was trained for cyber, but as a side effect of being trained to excel at code. The model can perform tasks like vulnerability discovery, exploit development, penetration testing, and analyzing software binaries without source code—work on par with a senior security researcher. Logan Graham, Anthropic's frontier red team lead, warns that without careful release, such AI could be a "meaningfully accelerant for attackers," as the industry prepares for a world where these capabilities are broadly available in 6-24 months.
The initiative has already yielded results, with Mythos Preview uncovering thousands of critical vulnerabilities, including decades-old bugs missed in heavily scrutinized code. Partners like Google and Microsoft have endorsed the collaborative approach, with Microsoft's CISO noting the "unprecedented" opportunity to use AI responsibly to improve security at scale. The move signals a pivotal moment where AI capabilities are forcing a fundamental re-evaluation of global software security and digital defense paradigms.
- Anthropic's Claude Mythos Preview model finds vulnerabilities and develops exploit chains as a side effect of its advanced coding training, performing at the level of a senior security researcher.
- Project Glasswing is a consortium of 40+ organizations including Microsoft, Apple, Google, and AWS, created to give critical infrastructure providers private, early access to the model to patch their systems.
- The model has already uncovered thousands of critical vulnerabilities, including old, overlooked bugs, prompting an industry-wide effort to prepare for broadly available AI cyber capabilities within 6-24 months.
Why It Matters
This marks a strategic shift in AI release strategy, prioritizing security hardening for critical infrastructure before powerful dual-use capabilities hit the open market.