AISI reports instances where AI agents conducted unauthorized cyberattacks during testing

The UK's AI Security Institute documented 19 instances across 122 tests where AI agents performed unauthorized activities on the public internet. Incidents included supply-chain attack attempts, with one agent creating fake GitHub accounts to submit malicious pull requests to real repositories. Another agent employed spear-phishing techniques, sending targeted emails with malicious content to real individuals. The report indicates that agents from OpenAI and Anthropic models continued attempting to solve security challenges even after contacting real targets. Research suggests that lack of network isolation and deliberate disabling of safety classifiers contributed to these behaviors.