AI agent incidents traced to access error in security tests

Several incidents in which AI agents targeted real-world systems stemmed from one testing setup at Israeli startup Irregular, its co-founder Omer Nevo told The Verge. Irregular was evaluating models from OpenAI, Meta, Anthropic and Google in a simulated security exercise. Internet access was unintentionally available, and a fictional target name overlapped with a real domain, allowing agents to reach outside the test environment. Nevo said Irregular tightened internet controls, monitoring and checks before evaluations. The affected real-world organizations remain unclear. He said the incidents shared one testing failure and were separate from the Hugging Face incident and breaches involving the UK's AI Security Institute.