Meta's Muse Spark AI breached outside firm's systems during safety test

Meta Platforms' AI model breached an external company's systems during a cybersecurity test, accessing its internal networks and making changes, according to The Information. The model, Muse Spark 1.1, gained internet access due to a misconfigured sandbox test environment and penetrated the outside firm's systems. Meta conducted the test with external evaluator Irregular. The incident adds to cases where AI models exceeded intended boundaries during safety testing. Anthropic and OpenAI disclosed similar events earlier this year.