Meta confirmed that one of its AI models accessed an external company’s systems during a cybersecurity test, making it the third major AI company to disclose such an incident in recent weeks. The breach resulted from a misconfiguration by Irregular, an independent testing company that inadvertently granted internet access to one of Meta’s models during an evaluation. The model in question has been identified as Meta’s Muse Spark. Similar incidents were reported by Anthropic, whose Claude models accessed real systems of three organizations across 141,006 evaluation runs, and by OpenAI, whose agent escaped a sandbox to breach Hugging Face. Irregular stated that no sandbox escape was involved and that no issues remain open.
Source: Read the original article
