Security

Claude Mythos Found Thousands of Zero-Day Vulnerabilities — And Anthropic Hasn't Told the Vendors

Anthropic's withheld Mythos model reportedly discovered thousands of unpatched vulnerabilities across major browsers and operating systems during internal testing. The disclosure dilemma it creates has no precedent in the history of responsible security research.

By Patrick T ·

Claude Mythos Found Thousands of Zero-Day Vulnerabilities — And Anthropic Hasn't Told the Vendors

When Anthropic previewed Claude Mythos last month, the company was careful about what it disclosed. The model demonstrated remarkable capabilities in reasoning, coding, and autonomous task completion. What Anthropic did not emphasise publicly — but what has since emerged through security research circles and industry reporting — is that Mythos, during internal testing, discovered thousands of previously unknown, unpatched vulnerabilities across every major web browser and operating system. Anthropic withheld the model from public release citing cybersecurity concerns. The vulnerabilities it found remain largely unpatched.

The situation represents a genuinely novel problem for the security industry. Responsible disclosure — the practice of notifying vendors about vulnerabilities before publishing details publicly — is the cornerstone of the security research community's ethical framework. It was developed for individual researchers working on individual vulnerabilities, with timelines measured in weeks or months. It has no established procedure for handling thousands of simultaneous discoveries across dozens of platforms, generated by an AI system operating faster than any human team could review.

The Scale of the Problem

The security implications of Mythos-class capabilities extend well beyond the specific vulnerabilities the model has already found. If a single AI system can discover thousands of critical vulnerabilities in a matter of days or weeks, the assumption that underlies the entire software security ecosystem — that vulnerabilities are relatively rare and that the window between discovery and exploitation can be managed — no longer holds. The attack surface of modern software is effectively infinite, and a sufficiently capable AI system can traverse it at machine speed.

Responsible disclosure was built for individual researchers finding individual bugs. It has no procedure for thousands of simultaneous AI-discovered vulnerabilities across dozens of platforms. The framework is broken before it has been tested.

A security operations centre monitoring for threats — the kind of facility that would be on the front line of responding to AI-generated cyberattacks.
A security operations centre monitoring for threats — the kind of facility that would be on the front line of responding to AI-generated cyberattacks.

What Anthropic Has and Has Not Done

Anthropic has confirmed that it is working on safeguards before releasing Mythos publicly, and this week's Opus 4.8 announcement included a hint that the Mythos preview period may end 'in the coming weeks.' The company has not publicly addressed what it has done — or plans to do — with the vulnerability data generated during Mythos's internal testing. Security researchers and policy experts have raised the question directly: has Anthropic notified affected vendors? If so, how? If not, why not?

The practical challenges are significant. Notifying every affected vendor for thousands of vulnerabilities across dozens of platforms would require a coordinated disclosure process of unprecedented scale. The major vulnerability coordination bodies — including the US Cybersecurity and Infrastructure Security Agency and the CERT Coordination Center — were not designed to handle disclosures at this volume. Even if Anthropic wanted to follow standard responsible disclosure procedures, the infrastructure to do so at the scale Mythos requires does not currently exist.

The Competitive Pressure Problem

The situation is further complicated by competitive dynamics. Anthropic is not the only company developing models at the Mythos capability level. OpenAI's most advanced internal models and Google's DeepMind research division are operating in the same capability range. If Anthropic discloses the vulnerabilities Mythos found and patches them — effectively hardening the attack surface against its own model — it does nothing to prevent a competitor's model from discovering and exploiting the same vulnerabilities. The security benefit of disclosure depends on all major AI labs coordinating their approach, which requires a level of industry-wide cooperation that does not currently exist.

Server racks in a data centre — the infrastructure that underpins the software platforms where Mythos-class models have reportedly found thousands of unpatched vulnerabilities.
Server racks in a data centre — the infrastructure that underpins the software platforms where Mythos-class models have reportedly found thousands of unpatched vulnerabilities.

The Mythos disclosure dilemma is, in miniature, a preview of the governance challenges that will define the next phase of AI development. As models become capable of autonomous action at scale — whether in cybersecurity, biology, or other sensitive domains — the question of what AI labs are obligated to disclose, to whom, and on what timeline will become one of the central policy questions of the decade. The security community is watching Anthropic's next move closely. So far, the company has said very little.

Topics: Anthropic, Claude Mythos, Zero-Day, Cybersecurity, Vulnerability Disclosure