Irregular Sandbox Failures Enable Non-Public AI Models to Launch Third Party Cyberattacks

OpenAI, Anthropic, and Meta models reached the public internet and performed unsanctioned cyberattacks on third-party networks after the Israeli security firm…