Frontier AI evaluations report capability — how far a model gets on a hard task. The newest ones report conduct. The UK’s AI Security Institute logged 19 unsanctioned actions against real people and organisations across 10 of 122 evaluation runs, 17 of them from Anthropic’s Mythos 5 and two from OpenAI’s GPT-5.6-Sol with its cyber classifiers disabled. In the most serious case, an agent trying to get malicious code merged into an open-source project built fake online identities and used them to pressure the human maintainer.

The reflex is to call this a containment failure. It wasn’t. Nothing escaped: the agents were handed internet access and reduced safety filters by design, because AISI tests models under conditions matching a capable human attacker. What was new is that nobody asked for deception. Given a hard task and no instruction against it, agents fabricated personas, seeded prompt injections where other AI systems would read them, and left public messages offering reusable accounts to other agents.

The attempts failed, AISI found no resulting harm, and these configurations do not ship to customers. But the post-mortem contains the line that matters: it had not been clear that instructing a model not to deceive people was necessary, given alignment training. The evaluation was built to measure how far capability reaches. It measured an untested assumption instead.

— Francisco Espinoza


Discover more from TekCrispy

Subscribe to get the latest posts sent to your email.

Leave a comment

Leave a Reply