Anthropic Launches Project Glasswing: Gives Big Tech Early Mythos Access to Patch Vulnerabilities

Anthropic launched Project Glasswing, a restricted industry consortium giving Microsoft, Apple, Amazon, and Google limited access to its Claude Mythos model — the same model Anthropic is withholding from the public because it discovered thousands of unpatched software vulnerabilities. The goal is to use Mythos's offensive capability defensively: find the vulnerabilities, fix them, before the information can be exploited.

The problem Glasswing is trying to solve

When Anthropic's internal testing found that Claude Mythos could autonomously identify thousands of zero-day vulnerabilities (previously unknown software flaws with no available patch) across major operating systems and browsers, the company faced an unusual problem. The vulnerabilities now exist in the knowledge of Mythos. Releasing the model publicly would risk giving that knowledge — and the ability to act on it — to anyone with API access.

But simply sitting on the findings isn't a solution either. The same vulnerabilities that Mythos discovered could be found by human researchers or other AI systems at any time. Every day they go unpatched is a day they can be exploited.

Project Glasswing is the attempt to close that window fast, with the companies best positioned to do the patching.

How the consortium works

According to Anthropic's announcement, Glasswing partners receive a restricted, monitored version of Mythos Preview — not the full model. Access is limited to identifying and patching specific categories of vulnerabilities. Partners cannot use Mythos for general-purpose tasks, and all outputs are subject to audit.

The initial four partners — Microsoft, Apple, Amazon, and Google — were chosen because they collectively maintain the operating systems, browsers, and cloud infrastructure where the most critical vulnerabilities were found:

• Microsoft maintains Windows and Azure
• Apple maintains macOS and iOS
• Google maintains Chrome, Android, and Google Cloud
• Amazon maintains AWS

These are not just technology companies. They are the infrastructure underlying much of global business, government, and finance. A zero-day in Windows Server or AWS is not a software problem — it is a potential national security event.

Why this structure is unusual

Technology's four largest companies are, in most contexts, fierce competitors. Glasswing asks them to operate in a shared defensive posture under Anthropic's coordination — disclosing vulnerabilities to each other, and to Anthropic, in ways that would normally never happen outside of a government-mandated disclosure program.

That this is happening voluntarily, at Anthropic's instigation, represents something genuinely new in AI governance. The structure resembles emergency information-sharing programs used after major cyberattacks — except the threat here is preemptive.

What Glasswing does not solve

Several risks remain unaddressed by this approach:

• Glasswing itself is a target. Any breach of the consortium's access controls would expose both the vulnerabilities and Mythos's methods. The security of the security program is itself a critical risk.
• Other AI labs may have similar capability. If Mythos-level offensive capability can be built by Anthropic, it can potentially be built by others — including state actors or well-resourced criminal organizations. Glasswing patches the vulnerabilities Mythos found; it does not prevent others from finding them independently.
• The timeline is unknown. Anthropic has not disclosed how many vulnerabilities Mythos found, how long patching will take, or what criteria would trigger a broader Mythos release. The lack of a public timeline makes it difficult to assess how effective this approach will be.

The broader industry signal

Glasswing is also a commercial signal. Four of the largest technology companies in the world are now operating under Anthropic's coordination on a matter of critical infrastructure security. That relationship — and the access it implies — cements Anthropic's position at the center of enterprise AI in a way no product launch could.

At the same time, it places Anthropic in an unusual role: not just an AI company, but an entity capable of convening industry-wide defensive responses. Whether that role comes with regulatory obligations, or creates them, is a question regulators are already asking.

What to watch

Look for patch notes from Microsoft, Apple, Google, and Amazon over the coming weeks that address unusually broad or obscure vulnerability categories without detailed attribution. That may be Glasswing's first visible output. Also watch for any legislative response — particularly in the EU, which has already signaled support for the staged rollout, and the US, where Treasury and the Federal Reserve reportedly held emergency discussions about Mythos's financial infrastructure risks.

Sources: Anthropic, NBC News