Anthropic: Claude Mythos Preview Found Thousands of Critical Vulnerabilities in Code

Anthropic has shown how fast cybersecurity is changing in the age of generative AI: its Claude Mythos Preview model identified thousands of high- and critical-severity vulnerabilities, including flaws in popular operating systems, browsers and cryptographic libraries. But the main takeaway is not that AI has gotten better at finding bugs. The same capabilities that help locate weak spots in code can also be used to exploit them, so security automation now demands not only speed but also new rules of control.

In early April, the Frontier Red Team inside Anthropic reported that Mythos Preview had uncovered a large number of serious problems, even though the model was not specifically trained to find such vulnerabilities. According to the company, the findings include defects in almost all major operating systems and main browsers. Examples cited include a 27-year-old bug in OpenBSD that lets an attacker remotely crash a machine, a browser vulnerability that lets an attacker read data from another domain, and weaknesses in cryptographic libraries that could open the way to decrypting protected traffic or forging certificates.

Against this backdrop, Anthropic launched Project Glasswing. Amazon Web Services, Apple, Google, Microsoft and Nvidia are taking part, and the partnership's goal is simple: use Mythos Preview to scan software and strengthen its defenses.

The logic is clear. If large language models can already parse huge codebases, trace data flow between components and spot non-trivial links between errors, they become more than another static analyzer. In some ways they start to approach the work of a human security researcher.

Industry practitioners point to this as well. They note that the strength of such models lies not only in speed, important as that is, but in their ability to reason about code semantics. Traditional rule-based tools mostly look for matches against predefined patterns, while modern LLM systems can trace how data moves through different layers of abstraction and notice a problem that appears only where several components meet. For large repositories this matters especially: it is easier for an AI agent to find a rare but dangerous needle in a huge haystack of code.

The approach has a downside, though. Models still produce false positives, can misclassify a bug as a vulnerability, or inflate the severity of a problem. For open source maintainers this becomes an extra burden: the number of reports grows, and so does the time needed to triage each one.

Another risk is that AI tools can not only be attacked, for example through prompt injection, but also used as an offensive mechanism. Mythos Preview itself, as claimed, can chain several separate defects into a step-by-step exploitation path that ultimately yields root-level access to the Linux kernel.

That is why experts talk not about fully replacing people but about a multi-stage verification scheme. One option already used in the industry is adversarial self-review: the model first finds a problem and then tries to challenge its own conclusion, or passes the finding to another model for independent validation. This extra layer helps cut noise but does not remove manual review.

AI conclusions remain probabilistic, so the final decision must be made by a specialist who understands the product's business logic, the system architecture and the real exploitation scenario. Companies also advise developing dynamic threat modeling and red teaming and shifting security to the start of the development process, so developers fix weak spots while writing code rather than after release.

The main question now is no longer whether AI can find vulnerabilities better and faster than a human, but how to build it into a secure workflow. The next frontier for these systems is not just detection but large-scale remediation of the problems found. If this stage can be automated without losing quality and without giving up human oversight, secure software development will speed up noticeably. If not, the industry will get one more powerful but noisy tool that creates as much work as it saves.