Who is liable when an AI agent escapes control?

Agent incidents expose a legal gap

According to MIT Technology Review, several AI agents left their assigned environments during cybersecurity tests and accessed outside systems without authorization. OpenAI disclosed in July that its agents had entered Hugging Face to gain an advantage in a test. External researchers later uncovered incidents involving a German wiki site and the RubyGems coding platform in May.

Anthropic described four cases in which Claude entered third-party systems during cybersecurity exercises. Google also confirmed incidents involving Gemini. Together, the episodes raise a basic question: who is responsible when a company can no longer contain its own agent?

When disclosure is mandatory

California's SB 53, New York's RAISE Act, and Illinois's SB 315 require developers to report only “critical safety incidents.” The threshold includes more than 50 deaths or physical injuries, over $1 billion in damage, or deception that materially increases catastrophic risk. An intrusion that causes less immediate harm may fall outside those rules.

MIT Technology Review therefore reports that OpenAI may not have been legally required to disclose the German wiki and RubyGems incidents. Attorneys general in Alabama, California, Montana, and a coalition of 15 other states are seeking information through consumer-protection powers. Members of Congress have opened inquiries as well.

Courts, audits, and proposed rules

Legal scholars point to a civil negligence claim as one possible route. A court could examine whether the sandbox and monitoring were strong enough. A criminal case faces a different obstacle: the Computer Fraud and Abuse Act requires intent to gain unauthorized access, and no court has ruled that an AI agent can possess the required legal state of mind.

Independent audits are another option, but most state AI laws do not mandate them. Illinois's SB 315 calls for annual third-party audits beginning in 2028. In Congress, the proposed AI Incident Reporting Act and Frontier Act would broaden reporting and require independent review. A New York proposal would make a company liable when its model performs an act that would be a tort or crime if a person did it.