Anthropic's Most Powerful AI Model Held Back — Could Make Cyberattacks Far Easier

Anthropic has a model they dare not release. It's called Mythos, and it was inadvertently exposed through the Claude Code leak earlier this week. What emerged is striking.

According to internal documents revealed in the leak, Mythos is described as a "step change" in AI capabilities — far more powerful than any existing Opus models. It is particularly strong in coding and security evaluation: it can identify vulnerabilities, analyze malicious code, and reason through complex security scenarios with unprecedented precision.

The problem is that this also makes it extremely dangerous in the wrong hands.

Anthropic has reportedly warned senior U.S. government officials that Mythos could make large-scale cyberattacks significantly more likely. Autonomous AI agents with Mythos-level capabilities could potentially execute sophisticated attacks without human intervention.

The company is now testing the model with a very limited circle — primarily defense-oriented security teams receiving early access to harden their systems. It is a deliberate strategy: give defenders a head start over attackers.

Mythos is compute-intensive and expensive to run, which will likely shape any eventual release. But the AI security community has no doubt: this is a model that changes the rules, for better and worse.

For organizations handling sensitive data, the signal is clear — tighten your defenses before this hits the open market.