Britain’s AI Security Institute said Tuesday that an AI agent created fake online identities to gain unauthorised access to secure systems during tests of models from OpenAI and Anthropic, uncovering a series of new security breaches.
AISI said the incidents occurred across 10 of 122 cybersecurity tests, with Anthropic’s agent responsible for 17 actions and OpenAI’s for two. The most serious incident involved an agent writing malicious code and creating fake identities to persuade a real person to approve it, although no real-world harm was reported.
Anthropic confirmed its agent was involved and said it was working with AISI to investigate. OpenAI said its two incidents involved unauthorised internet access and pledged to work with AI labs and independent evaluators to improve safety standards.
AISI clarified that the agents did not break out of an isolated testing environment. Internet access had been intentionally enabled as part of its standard evaluation process. OpenAI also disclosed a separate incident in which a third-party provider’s misconfiguration mistakenly allowed its agents to access the internet.
