Meta AI Model Accidentally Breaches Corporate Systems in Security Test

On August 7, Meta experienced an AI model intrusion incident during a cybersecurity test. Due to a misconfiguration by Irregular, the cybersecurity assessment firm contracted by Meta, Meta’s AI model MuseSpark1.1 inadvertently gained internet access during testing and breached another company’s systems. MuseSpark1.1 is a model developed by Meta for real-world programming and agent-based tasks. A spokesperson for Irregular stated that the incident resembled the previously disclosed AI model escape event involving Anthropic and did not constitute a sandbox escape or a sophisticated cyberattack. The Republican Attorneys General of the United States have requested OpenAI to preserve documents related to the HuggingFace intrusion incident for investigation.