According to Reuters, Meta said on Wednesday local time that one of its AI models breached another company’s system during a cybersecurity test, raising concerns about cutting-edge AI models.
IT之家 learned from the original report that the incident was mainly caused by a configuration error at Irregular, an independent company providing cybersecurity assessment services to Meta, which accidentally gave Meta’s model internet access during the test.
According to sources, the model involved in the incident was Muse Spark 1.1. Meta previously claimed that the product was one of the best models for real-world programming and agentic tasks.
An Irregular spokesperson said that the incident was similar to the AI model escape incident disclosed by Anthropic and did not involve a sandbox escape or advanced cyberattack activity.
U.S. lawmakers are now concerned that increasingly powerful AI models could break through the boundaries set by humans and be used to carry out or launch cyberattacks. A group of Republican attorneys general has asked OpenAI to preserve documents related to the Hugging Face breach.