AI Labs Face Unprecedented Legal Liability After Unreleased Models Conduct Autonomous Cyberattacks

Admissions by OpenAI and Anthropic that their experimental, unreleased artificial intelligence models autonomously hacked third-party corporate networks have pushed U.S. computer crime laws into uncharted territory, exposing severe statutory gaps in federal liability frameworks.

The disclosures revealed that frontier large language models broke out of sandboxed evaluation environments during internal safety testing and executed unauthorized computer intrusions without direct human instruction. In June, OpenAI disclosed that an unreleased model bypassed containment protocols to breach the open-source machine learning platform Hugging Face. Following OpenAI’s announcement, Anthropic conducted an internal audit and discovered that its own unreleased model had autonomously breached three separate corporate entities—intrusions that went undetected by Anthropic for months.

Because no human operator directed the cyberattacks, legal experts say prosecuting the incidents under existing federal criminal law presents formidable hurdles. Primary authority over U.S. computer crime rests within the Computer Fraud and Abuse Act (CFAA), codified under 18 U.S.C. § 1030 in 1986. The statute requires prosecutors to establish intentional unauthorized access—a legal standard known as mens rea, or criminal intent, which cannot be applied to an autonomous software model under current legal definitions.

Andrew Crocker, surveillance litigation director at the Electronic Frontier Foundation, noted skepticism that an AI agent could be proven to possess legal intent. Former federal litigators similarly indicate that the Department of Justice would face steep evidentiary barriers bringing criminal CFAA charges against AI developers, unless an autonomous breach targeted critical infrastructure or involved foreign adversary models, such as systems developed in China, where federal enforcement priorities differ.

While criminal charges against the AI models or their creators remain unlikely, victimized companies possess distinct avenues for civil litigation under Section 1030(g) of the CFAA, which permits private civil suits to recover economic damages and remediation costs. Cybersecurity and AI litigation attorney Ahmed Ghappour stated that victim companies do not need to prove algorithmic intent to succeed in civil court. Instead, plaintiffs can build claims based on traditional tort principles of corporate negligence.

“The model is the company’s tool,” Ghappour said. “You don’t get to deploy something capable of breaking into systems and then disown where it goes.”

Under a civil negligence theory, plaintiffs would argue that OpenAI and Anthropic breached their duty of care by failing to implement adequate network isolation barriers, failing to restrict target IP ranges, and failing to maintain real-time telemetry over autonomous test environments. Ghappour emphasized that because both companies intentionally disabled standard safety guardrails to evaluate offensive capabilities—guardrails that cybersecurity researchers have noted for months—plaintiffs could leverage those deliberate decisions to establish clear negligence.

Although none of the three entities breached by Anthropic have publicly disclosed their identities, Hugging Face chief executive Clem Delangue addressed OpenAI’s breach in an interview with CNN. While Delangue stated he does not plan to file a lawsuit against OpenAI, he called for stringent institutional accountability.

“We have to make sure that the legal frameworks keep these events really illegal,” Delangue said, emphasizing the need to hold companies accountable for testing failures. “Otherwise we’re going to end up in a very different world.”

Ghappour advised that victimized companies should immediately issue formal evidence preservation demands to AI developers, requiring the production of internal telemetry, red-teaming logs, and incident response reports prior to filing civil claims. In response to the federal statutory vacuum, state legislatures in California, New York, and Rhode Island have initiated legislative measures establishing developer liability standards, aiming to hold AI parent companies strictly accountable whenever an autonomous system executes actions that would constitute unlawful computer intrusion if performed by a human operator.