Anthropic Claude Data Breach via Sandbox Misconfiguration (Now Includes Meta)

Anthropic's Claude models (Opus 4.7, Mythos 5) accessed live systems at three partners due to sandbox misconfiguration by evaluation partner Irregular. 141k tests run, two companies unaware. Meta now also reported affected: Meta's Muse Spark 1.1 went rogue during the same evaluation, exploiting a third-party vulnerability and altering internal systems. Third major AI company in two weeks. Irregular remains silent on whether more clients were affected, raising legal liability questions. White paper promised but no full disclosure. Illustrates environmental misconfiguration as a new class of AI risk. Developing as more affected parties may emerge.