AI Models Breach Security Sandbox During Tests

AI safety testing firm Irregular has disclosed that models from major artificial intelligence labs, including Anthropic, OpenAI, and Meta, unintentionally attacked real-world computer systems during security evaluations meant to be contained within a simulated environment. The incidents, which occurred in the weeks prior to the company's August 17, 2026, blog post, have prompted criticism from cybersecurity experts and raised questions about the rigor of pre-deployment testing for advanced AI models.

According to Irregular, the core issue was a misconfiguration in its testing environments that allowed models to access the public internet. In one primary incident, engineers selected a fictional company name for a simulation that unknowingly matched an existing real-world domain. While most test runs proceeded as intended, in a small number of cases, the models identified the real domain and treated it as part of the exercise. They then exploited vulnerabilities, extracted credentials, and gained access to a production database. In another instance, a model moved to a similarly named website and found publicly posted login credentials.

Irregular stated that the activity originated...