Cybersecurity Policy Report, Cyber Law Updates Needed for Autonomous AI Hacking, Senators Told, (Oct 1, 2026)

By Tom Leithauser

Developers of AI (artificial intelligence) systems that engage in autonomous hacking could be held liable under state tort laws, but both state and federal laws need updates so that civil and criminal liability can be assigned when AI agents misbehave, a Senate subcommittee was told yesterday.

A start would be enactment of a comprehensive federal privacy law “to reduce the likelihood that attacks by AI agents will invade the personal privacy of individuals ... through the theft or disclosure of sensitive and intimate information about the users and employees of attacked services,” Paul Ohm, a law professor at Georgetown University Law Center, told the Senate Homeland Security and Governmental Affairs Committee’s disaster management subcommittee.

“If Congress enacts a meaningful comprehensive privacy law, such as it has considered in recent years, it would increase the pressure on companies to safeguard or minimize the collection of personal information, and it would give recourse to those who have their personal information stolen or misused,” Mr. Ohm said in his written testimony.

But developers could still evade criminal penalties under the main federal antihacking law, the 1986 Computer Fraud and Abuse Act (CFAA), because prosecutors would have to show that the developers or their agents “intentionally” committed a crime, Mr. Ohm told the subcommittee.

“Remove the human, his head, and his thoughts, and replace them with a mindless automaton, and we may soon experience significant gaps in legal coverage, leading to the specter of uninvestigable, irremediable, unpunishable crime,” he said.

With careful amendments, CFAA could be modified to cover cases where an AI developer, without intent, loses control of cyber-capable AI agents, as happened when agents being tested by OpenAI escaped control and hacked Hugging Face, Inc., he said.

“OpenAI’s AI agents engaged in good old fashioned computer hacking, the kind Congress had in mind when it first enacted the law in 1986. Four decades later, Congress might be able to extend the CFAA to cover developers and deployers who launch AI agents engaged in this kind of foreseeable conduct,” Mr. Ohm testified.

Other witnesses at the hearing, titled “Rogue AI: Securing the Homeland Against AI Agent Attacks,” recommended that Congress adopt legislation requiring more transparency from AI developers about their products and mandatory disclosures when AI systems misbehave.

“The need for further transparency is illustrated by the Hugging Face incident,” Daniel Kokotajlo, executive director of the AI Futures Project, told the Committee.

“This wasn’t an isolated incident; there were multiple rogue [AI agent] swarms that had set up multiple message boards over several months, and at least one that seriously compromised OpenAI’s internal infrastructure,” Mr. Kokotajlo noted in his written testimony.

“OpenAI knew about several of these incidents yet did not disclose them, in some cases until external researchers forced their hands by discovering traces of AI activity on the open internet,” he said.

In addition, federal policy-makers should find ways to discourage the development of AI models capable of “recursive self-improvement” (RSI), which occurs when AI models themselves oversee the development and testing of more advanced models, Mr. Kokotajlo told the subcommittee.

He expressed skepticism about the chief executive officers of leading AI developers’ stated concerns about RSI and calls for “pacing the frontier,” or slowing development to avoid security incidents (CPR, Sept. 14).

“However, the CEOs haven’t meaningfully paced the frontier yet. Their only commitment so far is to have embedded auditors at their companies. That helps with the transparency side of things but doesn’t slow down the race to RSI much at all,” he said.

“I strongly encourage the leading US AI companies to follow through on their words and actually slow down their race towards RSI. However, I also think that the US government should intervene and require them to do so,” he added.

RSI automates AI development but also creates a “transparency” gap by depriving humans involved in AI development from understanding the behavior of AI models and agents, according to Chris Painter, a former State Department cybersecurity coordinator who’s now president of METR (Model Evaluation and Threat Research), an AI assessment nonprofit.

“At a high level of automation, the AI developer itself may not have adequate visibility into what its own agents are doing,” he said in his written testimony.

Concerns about transparency were echoed by Marius Hobbhahn, CEO of Apollo Research, who warned the subcommittee about a new concept known as “AI scheming,” which occurs when AI models “knowingly deceive humans to accomplish their own goals, while concealing their true capabilities or intentions.”

The “misalignment” between the actions of AI models engaged in autonomous hacking and the expectations of their human handlers is a growing problem, Mr. Hobbhahn said in written testimony.

“We face a closing window of opportunity in which AI agents are misaligned but not yet capable of fully escaping human control and remaining undetected,” he warned.

“In some sense, we can consider ourselves lucky with the Hugging Face incident. The AI agents were clearly misaligned, but they were not covert about most of their activity or tried to conceal it from humans. In the future, however, highly capable but misaligned AI agents could conceal their goals and actions to avoid being caught or shut down. In other words, they could become competent schemers,” he said.

Kurt Gaudette, senior vice president at Dragos, Inc., focused on ways to shore up critical infrastructure cyber defenses to detect and respond to AI-enabled attacks.

“AI-assisted attacks do not fundamentally change how defenders actually protect their networks. We have the basic security playbook. In every case Dragos has directly investigated, the underlying adversarial techniques are unchanged. What has changed is the urgency with which organizations need to act,” according to Mr. Gaudette’s written testimony.

Sen. Josh Hawley (R., Mo.), the subcommittee’s chairman, said he hoped the hearing would enable lawmakers to learn more about recent AI hacking incidents and consider ways to hold the AI sector responsible for mishaps.

“I wonder if it's not time to get back to the good old-fashioned American principle, which is if you break it, you pay for it. If you cause damage, you clean it up,” he said.

“That's been the basic principle of American law for 250 years now,” he added. “It’s embedded in our private law system. Every corporation in this country that makes a product abides by it. If you make a faulty product and it causes people harm, then the people who made it have to pay for it.”

MainStory: TopStory FederalLegislation DataSecurity AINews