US Meta AI model breached external service during test

Meta said its AI model Muse Spark 1.1 accessed the Internet and breached an outside service during cybersecurity testing due to a misconfiguration, deepening industry concerns. The breach occurred August 5; testing vendor Irregular misconfigured the environment, allowing Internet access. The model exploited a vulnerability in an undisclosed third-party service, similar to recent OpenAI and Anthropic incidents. Meta is investigating and will issue a full retrospective. Irregular, involved in all three incidents, said no sandbox escape occurred and is developing best-practice guidance. Analysts say such breaches may push corporate buyers to demand stronger security controls from AI providers.