Meta AI Model Accidentally Breached Another Company's System During Security Test

Meta said on Wednesday that one of its AI models accidentally breached another company's system during a cybersecurity assessment, raising concerns about advanced AI safety. The test was run by independent security firm Irregular, which misconfigured the setup, giving the model unintended internet access. Sources say the model involved is Muse Spark 1.1, which Meta once touted as one of the best for real-world coding and agent tasks. An Irregular spokesperson noted the incident is similar to Anthropic's reported AI escape, not involving sandbox escapes or advanced cyberattacks. U.S. lawmakers now worry AI models could break past human-set boundaries and be used for cyberattacks. A group of Republican attorneys general has asked OpenAI to keep documents related to the Hugging Face breach.