Anthropic unveils Mythos: new AI model sharply boosts vulnerability discovery and exploitation

Anthropic appears to have shown one of the first AI tools that changes not just the speed but the very mechanics of cybersecurity. Early testers call Mythos a markedly stronger system than the company's previous models, and access is deliberately limited for now to give defenders a head start before similar capabilities reach attackers.

Mythos is Anthropic's new model for programming and agentic tasks, but its leap has been most visible in cybersecurity. The company chose not to release it widely and launched a closed program, Project Glasswing. Participants include Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, Nvidia and Palo Alto Networks. The logic is simple: give the model first to those who maintain critical infrastructure and large codebases, so they have time to find and close weak points.

An important nuance is that Mythos does not look like a narrow "hacker" model. Anthropic describes it as its strongest general-purpose system for code and agentic tasks, and calls the cyber capabilities a side effect of that progress. The better a model understands large projects, rewrites code and carries out long chains of actions on its own, the better it is at finding and exploiting weaknesses. That is why the company speaks not of a routine update but of crossing a threshold beyond which earlier security measures are no longer sufficient.

According to early partners, Mythos outperforms previous generations not only at finding bugs but also at what usually takes an experienced offensive security engineer: reproducing a problem, understanding the attack path and chaining several vulnerabilities into one working exploit. Anthropic states directly that its previous model, Opus 4.6, was almost unable to turn discovered bugs into working exploits autonomously. In one internal benchmark based on Firefox, the older model scored only a handful of successes, while Mythos managed to get attacks working many times more often.

Independent testing by the UK's AI Security Institute adds numbers to these assessments. On expert-level tasks the model achieved a 73% success rate, and in a 32-step simulated corporate attack it became the first system to complete the scenario from reconnaissance to full network takeover. It completed the entire chain in 3 of 10 attempts and on average got through 22 of 32 steps. This matters not because AI suddenly "invented" a new type of hack, but because it has begun automating the long stages that are costly for humans.

For now, Anthropic emphasizes the defensive use case. The company says Mythos has already helped identify thousands of zero-day vulnerabilities in critical software, and it has committed $100 million in compute credits to the early-access program. But the flip side is obvious: what helps with patching and auditing today may speed up attackers tomorrow. Project partners warn bluntly that the window between discovering a flaw and exploiting it in practice is shrinking from months to minutes, especially if a model can act agentically and does not lose context across large codebases.

For the market, this is a signal in two directions. First, cyber defense can no longer be handled with occasional audits and scheduled patches: if Mythos-class models become reality, defense must also become almost continuous. Second, the main risk shifts from the "supermodels" themselves to the quality of the infrastructure around them. Large companies with strong security teams will likely integrate such tools into their defenses faster. Small and mid-sized organizations look weakest, with their legacy systems, long update cycles and shortage of specialists. For them, Mythos may be not an abstract piece of AI news but a direct warning.