45 organizations working on cybersecurity together is already unusual. Getting Apple and Google in the same coalition is rarer still. That's the structure Anthropic has assembled with Project Glasswing, a new initiative using Claude Mythos Preview - a model Anthropic built specifically for security research - to test AI-powered defenses across critical software.
What Claude Mythos Preview Is Designed For
Claude Mythos Preview is oriented toward the kind of work security researchers actually do: reasoning through complex attack chains, identifying vulnerabilities in code before they're exploited, and suggesting patches that address root causes rather than surface symptoms. This is categorically different from asking a general-purpose AI assistant to review code for bugs. Security-focused models need detailed knowledge of how exploits work, which raises obvious concerns about the same capability being misused offensively.
Anthropic's response to that tension is the coalition structure: Claude Mythos Preview access is tied to Project Glasswing participation rather than available as an open API. That limits access, but also limits what can be learned from broader real-world deployment.
Scale and Coordination
The breadth of the coalition matters for a specific reason. Critical software vulnerabilities rarely live in isolation. They span operating systems maintained by one company, cloud platforms operated by another, and enterprise applications built on both. Getting Apple and Google into a coordinated defensive initiative is genuinely difficult to organize. The fact that 45+ organizations signed on suggests the industry has reached a shared conclusion that AI-powered attack capabilities are developing faster than individual organizations can defend against on their own.
If Project Glasswing works as intended, the signal to watch for is coordinated vulnerability disclosures and patch releases across participating organizations over the next 12 to 18 months - not research papers, but actual software updates addressing issues Claude Mythos Preview surfaced. Apple and Google's security update cadences would be particularly telling. An acceleration in either, combined with coordinated disclosure timing, would indicate the coalition is producing real results.
For security teams at organizations outside the coalition, the practical implication is this: if AI can surface vulnerabilities at the pace Anthropic is claiming, expect the rate of disclosures for widely-used software to increase meaningfully over the next year. Organizations that are already slow to deploy security patches will find themselves in a worse position as AI-accelerated discovery outpaces their response cycles. The harder question - whether a model capable of finding critical vulnerabilities can be kept from those who would exploit them - doesn't have a clean answer here. Project Glasswing is a reasonable structural bet. Whether the coalition maintains the discipline to keep it that way is less certain.