Anthropic Introduced Mythos: New AI Model Sharply Enhanced Vulnerability Detection and Exploitation

Anthropic appears to have unveiled one of the first AI tools that changes not just the speed but the very mechanics of cybersecurity: early testers call Mythos a noticeably more powerful system than the company's previous models, and access to it is intentionally limited for now to give defenders a head start before similar capabilities reach threat actors. Mythos is Anthropic's new model for programming and agent-based tasks, but it's specifically in cybersecurity where the leap has proven most significant. The company decided not to release it for broad access and instead launched the closed Project Glasswing program.

It involves Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. The logic is straightforward: first give the model to those maintaining critical infrastructure and large codebases so they have time to find and close vulnerabilities. An important nuance is that Mythos does not look like a narrow "hacker" model.

Anthropic describes it as the most powerful general-purpose system for code and agent tasks, and calls the cybersecurity capabilities a side effect of this growth. The better a model understands large projects, rewrites code, and autonomously executes long chains of actions, the better it finds and exploits weaknesses. This is precisely why the company speaks not of a local update but of crossing a threshold after which previous security measures are no longer sufficient.

According to feedback from early partners, Mythos handles not just vulnerability discovery better than previous generations, but also tasks that typically require experienced offensive security engineers: reproducing the problem, understanding the attack path, and chaining multiple vulnerabilities into a working exploit chain. Anthropic explicitly states that its previous Opus 4.6 was almost incapable of autonomously bringing discovered errors to a working exploit.

In one internal benchmark based on Firefox, the old model achieved only a handful of successful runs, whereas Mythos could repeatedly develop attacks to a functional state. An independent assessment by the British AI Security Institute adds numbers to these evaluations. On expert-level tasks, the model showed 73% success, and in a simulated corporate attack with 32 steps, it became the first system to successfully execute the scenario from reconnaissance to full network compromise.

It managed to complete the full chain in 3 out of 10 attempts, and on average it progressed through 22 of the 32 steps. This matters not because AI suddenly "invented" a new type of breach, but because it has begun automating long and expensive stages for humans. For now, Anthropic emphasizes the defensive use case.

The company states that Mythos has already helped identify thousands of zero-day vulnerabilities in critical software, and has allocated $100 million in compute credits to the early access program. But the flipside is obvious: what helps with patching and auditing today could accelerate attacker operations tomorrow. Project partners warn explicitly that the window between vulnerability discovery and practical exploitation is shrinking from months to minutes, especially if the model can act as an agent and doesn't lose context on large codebases.

For the market, this is a signal pointing in two directions. First, cyber defense stops being a process that can be handled with infrequent audits and scheduled patches: if models at Mythos's level become reality, defense too must become nearly continuous. Second, the primary risk shifts away from the "super-models" themselves toward the quality of the infrastructure around them.

Large companies with strong security teams will likely integrate such tools into their defenses faster. Worst off are small and medium-sized organizations with accumulated legacy systems, long update cycles, and a shortage of specialists. For them, Mythos's arrival may be not an abstract AI world news item but a direct warning.