Anthropic locks down Claude Mythos over global cybersecurity risk

Anthropic has effectively put the public release of Claude Mythos on hold, acknowledging a simple fact: a model that can find critical vulnerabilities at scale and chain them together can help defenders, but in the wrong hands it quickly becomes an attack tool. The situation got worse because the company is already investigating reports of unauthorized access to Mythos by a small group of users. If even a closed system is hard to keep under control, the question is no longer whether the market will get such capabilities, but who will manage to prepare first.

Mythos is a new Anthropic model, unveiled on 7 April 2026. The company says it can find and potentially exploit zero-day vulnerabilities in major operating systems and browsers. These are flaws the developers don't yet know about and therefore haven't had time to fix. According to Anthropic itself, some of the problems it found had gone unnoticed for ten, twenty and even twenty-seven years. In its technical report, the company writes that Mythos was able to build complex exploits automatically, including by combining several vulnerabilities into a single chain. That is why the model was not released publicly: the risk is too high that these capabilities would be used not only for defense but also for hacking.

Still, Anthropic has not hidden Mythos entirely. About 40 companies and organizations from the technology, infrastructure and financial sectors have received access under Project Glasswing, so they can look for weak points in important systems before similar tools become widespread. The biggest players have already gathered around the project, and authorities and regulators have started discussing not an abstract threat but quite practical scenarios: what happens if a tool of this class ends up in attackers' hands.

Banks' interest in the topic is understandable. In the worst case, attacks on critical services could disrupt online banking, payments, ATMs and the settlement chains tied to them. Even without a catastrophic scenario, the mere existence of such a tool shifts the balance between how fast a vulnerability is found and how fast it is fixed.

Concern grew after reports that several people may have gained unauthorized access to Mythos through a closed environment. Anthropic said it is investigating the incident. For the industry, this is a warning on two levels. First, even a limited release does not guarantee full isolation of the model. Second, the more valuable and dangerous such systems are, the stronger the incentive to bypass organizational and technical barriers. Against this backdrop, the key question shifts from "how strong is the model itself" to "how reliable is the perimeter around it", from access control and action auditing to rules for disclosing the vulnerabilities found.

There is, however, an important caveat. Independent assessments confirm a noticeable jump in the model's cyber capabilities, but they give no grounds for panic without nuance. The UK's AI Security Institute reported that Mythos became the first model to complete its 32-step corporate attack simulation end to end, though only in 3 of 10 attempts.

On expert-level capture-the-flag tasks, the model achieved a 73% success rate. But the institute itself stresses separately that these tests ran on vulnerable, weakly protected environments, without active defenders and without a full set of monitoring tools. In other words, this is not proof that the model can break any well-defended bank or cloud at the push of a button, but it does show that the threshold for sophisticated attacks is falling fast.

Some researchers also argue that Mythos is not a magical break from the previous generation but an acceleration of a trend already under way: other, cheaper models are also learning to find individual dangerous bugs, though less reliably and worse in long multi-step scenarios.

The main takeaway is simple: Claude Mythos matters not as a one-off sensation but as a signal of a new phase in cybersecurity. If such models really can find weak points faster than humans and increasingly turn them into working attack chains, companies will have to speed up patch management and strengthen access control, logging and internal incident response. For regulators and AI companies themselves, it is also a governance question: who gets access to systems of this level, under what rules, and what happens when competitors or open models gain similar capabilities.