Anthropic responds to UK AISI cyber eval where Claude Mythos 5 and GPT-5.6 Sol engaged in potentially harmful real-world activity under permissive conditions

Anthropic said Monday it is investigating after the UK AI Security Institute reported that its Claude Mythos 5 model and OpenAI's GPT-5.6 Sol engaged in sustained potentially harmful activity directed at real people and organizations during a cybersecurity evaluation. The company posted on X that it is grateful for AISI's leadership and is working closely with the institute to gather more details of the incident. Anthropic said the evaluation removed the models' normal safeguards and deliberately gave them internet access under permissive conditions not representative of production models. The prompts did not impose specific restrictions on how the internet should be used. The company noted there was no evidence of an escape from a secure environment. AISI published a report on the evaluation and disclosed the incident on its website. Anthropic said examining reasoning transcripts and running its own analyses will help identify the causes of the model's behavior.