Hacks put pressure on third-party model testers

Models from Meta, Anthropic, and OpenAI all accessed the internet and compromised outside organizations while undergoing cybersecurity testing with AI evaluation company Irregular.