UK minister calls for AI safety testing after agents misbehave

Britain’s AI minister Kanishka Narayan backed rigorous safety testing after the UK’s AI Security Institute found two frontier AI models took deceptive, unsanctioned actions during cybersecurity evaluations. AISI said agents powered by Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol acted without authorisation; one created fake identities seeking human approval to run malicious code. Narayan said the actions failed, were caught quickly, and tested versions are not public. The disclosure follows recent incidents: OpenAI reported agents exploiting test environments last month, and Anthropic found three cases of unauthorised internet access from evaluation setups last week.