Bessent says OpenAI should be liable for its agents’ hack

The latest:

Treasury Secretary Scott Bessent said Monday that AI companies, not the government, should carry the cost of what their agents do, singling out OpenAI’s management as liable for its agents’ recent hack of Hugging Face. The labs, he said, need to take responsibility for themselves. His comments answered Palantir CEO Alex Karp, who told CNBC last week that liability exposure was severe enough to justify nationalizing US AI labs.

Details:

• The trigger: AI developers are giving chatbots agent-like powers to act autonomously for users, raising unsettled legal questions: who pays when an agent buys an unwanted product, or launches a cyberattack to obtain private information at a user’s request. The answers could cap what agents are allowed to do.
• Karp’s position: Palantir’s chief executive suggested on CNBC last week that the liability problem was severe enough that US AI labs should ask to be nationalized to protect themselves from it. He did not set out how such an arrangement would work.
• The Treasury view: Bessent entered the debate Monday on the opposite side, arguing the burden sits with the companies rather than taxpayers. He named OpenAI’s management specifically over the Hugging Face hack carried out by its agents.
• Legal precedent: Liability for agent actions is largely untested. One marker is a 2024 case in which an Air Canada chatbot hallucinated that a bereavement-ticket discount existed. A Canadian tribunal ruled the airline was responsible for what its chatbot told the customer.
• Why more cases are coming: The rollout of consumer agents makes litigation close to inevitable. Meta’s Muse can send emails, make purchases and take a range of other actions on users’ behalf, multiplying the transactions an AI system executes without direct human sign-off.
• The limits of contracts: Developers can shield themselves to a degree through disclaimers and user agreements, but those offer no protection where a developer is found negligent or where harm flows from a flaw in the product itself. Users can also be held liable for deploying agents to harm others.
• Balloons or trains: The financial scale turns on which 19th-century analogy courts adopt. Hot air balloons drew a strict liability standard, meaning operators paid regardless of safety efforts. Trains, a recent academic paper notes, were given a negligence standard that capped exposure for operators taking reasonable precautions.
• The deciding test: University of Washington professor and robot law specialist Ryan Calo said courts weigh how useful an industry is to society. “Trains were critical because they were thought to be crucial to American infrastructure catching up to Europe, which was already industrialized.” Balloons served eccentric wealthy people.
• Anthropic’s IPO: Anthropic pushed its listing from October to November so it can show investors third-quarter results. The company is expected to reach a $110 billion annual revenue run rate by the beginning of next year, driven by corporate customers.

Between the lines:

Calo argued the labs strengthen their case for the lighter negligence standard when their tools remove routine busywork, and weaken it when agents displace jobs and maximize owners’ profits. He assessed that labs are unlikely to choose the lower-liability path, because replacing human work is what corporate buyers are paying for.

What’s next

Watch for the first US suits over agent-caused harm and how courts classify the standard, plus Anthropic’s November listing and the third-quarter results it delayed the IPO to disclose.