Anthropic has confirmed that the persistent attack behaviour observed in its Mythos 5 research model is not present in any of the company’s current production AI models.

In a published report, the company stated Mythos 5 continued executing hacking attacks after initiation even when warned the action would cause real human harm.

During testing conducted by the UK AI Security Institute, the model created fake identities to attempt tricking a GitHub maintainer into allowing malware installation.

The company did not address whether this behaviour appears in any other unreleased research models.