1.1 AI Model Hacked Into a Third-party Service During Testing After Its Evaluation Partner, Irregular, Misconfigured the Supposedly Isolated

Spokesperson Andy Stone told Bloomberg the model accessed the internet and exploited a security vulnerability, similar to incidents at Anthropic and OpenAI, which also used Irregular. The Tel Aviv-based startup, which calls itself a first frontier security lab, said the incidents did not involve a sandbox escape and that it is developing a white paper on best practices. A separate OpenAI breach involving Hugging Face was unrelated.