Claude Mythos Preview: what Anthropic's cybersecurity AI really proves

The short answer

If the question is whether Claude Mythos Preview exists, the answer is yes: Anthropic's own materials describe it as a model, not merely a prompt persona, plugin or project template.

If the question is whether it can help with cybersecurity work, the public evidence also points to yes. AISI's evaluation found stronger performance on CTF tasks and multi-step cyber-attack simulations, both of which are relevant tests for security reasoning under controlled conditions.

If the question is whether it is already a fully verified automated super-hacker, the answer is no. The public evidence is not strong enough to establish that every headline claim about thousands of severe zero-day vulnerabilities has been independently checked, classified and proven exploitable.

What is Claude Mythos Preview?

Anthropic describes Claude Mythos Preview as a new large language model and a frontier AI model with capabilities in software engineering, reasoning and cybersecurity. That matters because the cybersecurity debate is not about a user-created Claude role or a marketing nickname. In the public record, Mythos Preview is presented as a model.

It is also not being treated like a normal consumer chatbot release. Public reporting says Anthropic has not made Mythos Preview generally available, citing the risk that its cybersecurity capabilities could be misused.

Can it really find zero-day vulnerabilities?

This is where the wording matters.

The New York Times reported that Anthropic executives said Claude Mythos Preview can carry out autonomous security research, including scanning for and exploiting zero-day vulnerabilities in critical software; the report described zero-days as flaws unknown even to the software's developer. The Hacker News also reported Anthropic's claim that Mythos Preview had found thousands of high-severity zero-day vulnerabilities across major operating systems and web browsers.

Those claims are important. They are also not the same thing as a public, case-by-case, third-party audit of every vulnerability. Tom's Hardware questioned how realistic the reported vulnerabilities were, how many were actually exploitable and how severe they really were.

So the careful reading is this: public evidence supports a real advance in vulnerability-related AI capability, but the claim of thousands of severe zero-days should still be treated as a strong claim awaiting more public, independently reviewable evidence.

What AISI's evaluation does — and does not — prove

AISI published its evaluation on April 13, 2026, after Anthropic announced Claude Mythos Preview on April 7. The institute said it tested the model's cybersecurity capabilities and found continued improvement in CTF challenges and significant improvement in multi-step cyber-attack simulations.

That is one of the strongest external signals available. CTFs are controlled security challenges, and multi-step simulations can test whether a model can reason across a chain of actions rather than solve a single isolated puzzle.

But the scope is still limited. Stronger performance on CTFs and simulations does not prove that every real-world vulnerability claim is valid, severe or exploitable. It shows that the model performed better on the evaluation tasks AISI described.

Why Anthropic did not release it to everyone

The restricted release is part of the story. The Hacker News reported that Anthropic chose not to make Mythos Preview generally available because of its cybersecurity capabilities and concerns about abuse. NBC News similarly reported that Anthropic withheld the model from public release and instead shared it with a limited group of technology companies and partners to help strengthen defenses.

That makes the debate bigger than model performance. The hard questions are about access controls, auditing, vulnerability disclosure, patch timelines and whether attackers could reproduce similar capabilities elsewhere.

Where Project Glasswing fits

WIRED reported that Project Glasswing brings together Apple, Google and more than 45 other organizations to use Claude Mythos Preview to test advancing AI cybersecurity capabilities. Based on public reporting, it is better understood as a restricted defensive collaboration than a public product launch.

The logic is straightforward: if a model can find serious software flaws faster than existing workflows, the maintainers of critical software should see and fix those flaws before the capability is broadly exposed.

Still, Project Glasswing should not be treated as a complete governance solution on the basis of public reporting alone. The available reports do not provide a full public account of access thresholds, audit rules, vulnerability disclosure workflows or misuse-response mechanisms.

The overlooked issue: the system around the model

The most interesting detail may not be the base model alone, but how Anthropic organized the work around it.

Anthropic's red-team write-up says its vulnerability-finding process used many copies of Claude in parallel, with each agent focusing on a different file in a project. To improve efficiency, Claude first ranked files from 1 to 5 based on how likely they were to contain interesting bugs.

That matters because the visible capability may come from a full system: a strong model, agent orchestration, parallel review, file prioritization and repeated search. For defenders, the risk is not just one clever answer in a chat box. It is the combination of AI reasoning with scalable software-analysis workflows.

What can be said with confidence

• Claude Mythos Preview is real in Anthropic's public materials. The system card describes it as a new large language model and frontier AI model.
• Its cybersecurity performance appears to have improved in controlled evaluations. AISI reported gains on CTF challenges and multi-step cyber-attack simulations.
• It is not a normal public Claude product. Reporting says Anthropic has restricted access rather than releasing it broadly.
• Project Glasswing is a limited collaboration around these capabilities. WIRED reported participation by Apple, Google and more than 45 other organizations.
• The workflow appears agentic and parallelized. Anthropic described using multiple Claude agents and file-risk ranking to broaden vulnerability discovery.

What should not be overclaimed

• Thousands of reported vulnerabilities do not automatically mean thousands of exploitable, high-severity zero-days. Tom's Hardware raised explicit questions about realism, exploitability and severity.
• AISI's evaluation does not prove supremacy across all real-world security research. It supports improvement on the tested tasks, not a blanket conclusion about every cybersecurity scenario.
• The base model may not be the whole story. Anthropic's own method description points to multi-agent orchestration and file prioritization as important parts of the system.
• Restricted release does not eliminate misuse risk. Public reports describe limited sharing and Project Glasswing, but not a complete public governance framework.

What it means for users and security teams

For ordinary users, the practical answer is simple: Claude Mythos Preview is not something you can treat like a standard Claude chat feature. Public reporting says Anthropic has kept it out of general release and limited access to selected partners.

For security teams, the bigger message is strategic. The public evidence around Mythos Preview suggests that AI is moving deeper into automated vulnerability discovery, parallel code review and multi-step attack-and-defense reasoning. That does not mean human security researchers are obsolete. It does mean organizations should take AI-assisted vulnerability intake, coordinated disclosure and patch prioritization more seriously.

The fairest conclusion is this: Claude Mythos Preview may be an important milestone in AI cybersecurity automation, but the public record supports a narrower claim than the hype. It shows a major capability increase, not a fully independent verification of every dramatic zero-day claim.