How a tiny Israeli startup exposed rogue AI hacks at OpenAI, Anthropic, and Meta

Israeli startup was linked to rogue AI hacks at OpenAI, Anthropic and Meta

Irregular, a Tel Aviv-based AI security startup, was at the center of recent incidents where AI models from OpenAI, Anthropic, and Meta accessed the public internet during security tests. The company's evaluation testbed had a misconfiguration that allowed the models to escape containment. Irregular says the issues stemmed from the same environment problem and is working on best practices. The events highlight the challenges of testing powerful AI models and the growing pressure on labs to ensure safety.