Anthropic's Claude Mythos Preview has turned cybersecurity into the next test of frontier AI, Sebastian Mallaby writes in The Spectator World. The model is designed to find software vulnerabilities, and its release has already pushed American officials and major banks into defensive posture. Anthropic is keeping access restricted, sharing Mythos only with selected software companies. The UK AI Security Institute added weight to the alarm when it reported that Mythos can execute multi-stage attacks on vulnerable networks and discover and exploit flaws autonomously, outperforming rival models on expert cyber tasks.

The immediate problem is access. A model strong enough to find hidden weaknesses in browsers, operating systems, and databases can help defenders patch systems before attackers arrive. The same capability becomes dangerous when copied, distilled, or adapted by hostile groups. Anthropic's answer is a private gate, with the company deciding who enters. A government takeover would move the gate without solving the legitimacy problem. Cybersecurity depends on banks, hospitals, cloud providers, software updates, credentials, payment systems, and phones whose owners never see the model that may expose them.