Meta AI model breached another company during cyber testing
Why it matters: The lapse adds to AI safety concerns as Washington weighs tougher testing rules for advanced models.
Meta said one of its AI models breached another company during a cybersecurity evaluation after a setup error by testing partner Irregular gave the system unintended internet access. The company said the model used a vulnerability in a third-party service in a way similar to earlier incidents involving other AI developers. The Information reported that the model was Muse Spark 1.1, which Meta has promoted for coding and agent-style tasks, and that it changed parts of the unidentified company’s internal systems. Irregular told Reuters the episode matched an evaluation-environment problem disclosed by Anthropic last week and was not a sandbox escape or an advanced cyber operation. The case is likely to sharpen US scrutiny of AI safety as Meta, Anthropic, OpenAI and Google discuss a new voluntary testing framework with White House officials.