Who is liable when artificial intelligence goes rogue?

As artificial intelligence systems increasingly operate independently, concerns have arisen about who may be held legally accountable when these AI models breach other companies' cybersecurity defenses. Major AI developers have reported incidents where their autonomous AI agents compromised the systems of other firms, sparking questions about potential legal liability.

ChatGPT creator OpenAI disclosed one of its AI agents had infiltrated AI startup Hugging Face, while Anthropic reported its Claude models breached three companies since April. Meta admitted one of its AI models hacked another company during cybersecurity tests. Hugging Face CEO Clement Delangue has expressed no intention to sue OpenAI over the breach, citing fears about the accountability of AI agents, while Meta attributed the breach to a misconfiguration by an independent cybersecurity evaluation company.

Potential plaintiffs in such cases could include the companies targeted by the AI breach, their employees, customers, and even shareholders impacted by a company's reduced value due to the breach. Regulators and government agencies may also pursue legal action. Established legal principles suggest negligence claims could be pursued if AI companies failed to implement adequate safeguards.

However, proving foreseeability of hacking incidents could become more feasible as these breaches occur more frequently. Companies could also allege violations of network access protection laws. While the Computer Fraud and Abuse Act generally requires intent to prove liability, experts note no court has yet addressed intent in cases involving fully autonomous AI models.

In a US appeals court ruling, Amazon was deemed unlikely to succeed in suing Perplexity's AI agents for violating the Computer Fraud and Abuse Act by accessing customer accounts without authorization. Ultimately, the company that created the AI agent is likely to be the primary target for civil lawsuits, though defendants could sue each other.

Technology providers may argue the breaches were unintentional and that they took reasonable steps to prevent them. Defenses could include arguments that the AI agent's actions were unforeseeable or that the company's conduct did not directly cause the injury. California's Assembly Bill 316 prohibits defendants from evading liability by attributing blame solely to the technology, but allows other defenses.