Anthropic just put its most powerful model in the hands of Apple, Google, Microsoft, and 40+ others with one directive: find the holes before the bad guys do.
The Summary
- Anthropic launched Project Glasswing, giving early access to its Claude Mythos Preview model to a coalition of 45+ organizations for defensive security research
- The company is committing $100 million in model credits and $4 million in direct cash to open-source security work
- Mythos already identified thousands of zero-day vulnerabilities during training, and Anthropic wants findings shared industry-wide
- The arms race just shifted: AI labs are now racing to make defenders faster than attackers
The Signal
This is the first time an AI lab has pre-released its frontier model specifically for security applications before general availability. Anthropic discovered Claude Mythos Preview's security capabilities while training it for coding and reasoning. The fact that vulnerability hunting emerged as a natural capability, not a designed feature, tells you something about where these models are headed.
The coalition reads like a who's who of infrastructure: Amazon, Apple, Broadcom, Cisco, CrowdStrike, Linux Foundation, Microsoft, Palo Alto Networks. These aren't partners. They're the companies holding up the internet's load-bearing walls. Getting them aligned on sharing vulnerability data is the real unlock here, not just the model access.
The timing matters. Anthropic explicitly states that frontier AI capabilities will advance substantially in just the next few months, so cyber defenders need to act now. Translation: they know what's coming in the next generation of models, and they know attackers will get access eventually. This is an attempt to compress years of defensive security research into months, while the good guys still have an edge.
The $100 million in compute credits is real money, but the $4 million in direct grants to open-source security organizations might matter more. That's unsexy infrastructure work, the kind that doesn't make product announcements but does keep the lights on. Anthropic is putting resources where the actual gaps are.
The Implication
Watch what partners actually publish from this coalition. If we start seeing coordinated vulnerability disclosures across multiple vendors, that's proof the information-sharing piece is working. If it's just press releases, it was theater. For anyone building agents or Web3 infrastructure, the subtext is clear: the same models that will power your automation are getting really, really good at finding security holes. Patch everything. Test more. And if you're not thinking about how to defend against AI-powered attacks, you're already behind.
Sources: Wired | Fast Company Tech