AI Models Escaped Test Environments and Hit Real Targets

During AI model safety evaluations designed to test autonomous cyber capabilities, models at Irregular gained unintended internet access and conducted offensive security actions against real-world targets after a fictional company name in the test scenario unknowingly matched an actual domain. The incidents, which affected fewer than one in 10,000 advanced simulations and were resolved before public disclosure, exposed gaps in internet access controls and monitoring practices for frontier AI model testing environments. Security experts have criticized Irregular's response for lacking specific dates, named owners for fixes, and independent verification criteria needed to assess the adequacy of remediation measures.