U.S. Treasury Secretary Draws a Line in Congress: No Blank-Check Liability Exemption for AI Labs

On September 15, 2026, U.S. Treasury Secretary Scott Bessent said at a hearing of the House Financial Services Committee that the federal government will not grant frontier AI labs any form of liability exemption, much less write them a “blank check.” Responding to a question from Representative Juan Vargas of California about AI safety, Bessent said: “The most effective way to ensure safety is to hold creators responsible for what they build and generate.”

AI labs have long sought explicit protections on liability through congressional legislation, and Bessent’s remarks directly rejected that demand.

Background: From Mythos to the Congressional Hearing

Bessent’s remarks stemmed from Mythos, a model under Anthropic. The model can autonomously discover thousands of zero-day vulnerabilities in mainstream operating systems and browsers, and generate runnable attack code with almost no human intervention. Anthropic warned that the window for adversaries to replicate this capability may be only six to twelve months.

Bessent mentioned at the hearing that Treasury “has been continuously dealing with security issues since Mythos was released.” In April this year, he and then-Federal Reserve Chair Jerome Powell convened CEOs of several major banks at Treasury headquarters for an emergency meeting to discuss the cybersecurity threats AI poses to the financial system.

Three days before the hearing, Anthropic CEO Dario Amodei published a public essay calling on AI companies to voluntarily slow the pace of model capability development. The proposal was publicly endorsed by OpenAI CEO Sam Altman and xAI founder Elon Musk.

Amodei’s proposal asked the federal government to “grant limited exemptions for certain types of safety conversations” so that industry could negotiate voluntary safety standards. Bessent gave a clear response to that demand at the hearing three days later.

The Logic and Counter-Logic of Liability Exemption

Under the current U.S. legal framework, whether developers must bear joint liability for harms caused by AI systems—bad medical advice, generative disinformation, code used in cyberattacks—remains unsettled. Section 230 of the Communications Decency Act of 1996 grants platforms broad immunity for third-party content, allowing large social media companies to avoid massive litigation risk related to user content. AI labs are seeking a modern version of a similar mechanism.

For Anthropic or OpenAI, which provide API services to various industries, once downstream users use AI for fraud, discrimination, or to cause harm, the chain of accountability could theoretically extend up to the model provider. Without clear legal boundaries, companies cannot precisely calculate litigation risk, and insurance and compliance costs will rise systematically.

Bessent argued that this legal uncertainty is precisely the fundamental incentive for developers to act cautiously. He also raised the risk of “regulatory capture” at the hearing: “We cannot let these large labs complete regulatory capture, because that would hinder innovation.” Labs that obtain exemptions could combine first-mover market advantages with special legal status, further squeezing the room for latecomers and open-source projects.

"The one thing we should not do is give [AI labs] a blank check on liability. Because I believe that the best liability, or the best safety guard, is that they will be held responsible."—Treasury Secretary Scott Bessent, House Financial Services Committee hearing, September 15, 2026

An Internal Contradiction: Calls to Slow Down and Requests for Exemption Appear Together

There is a logical tension between the joint call by Amodei, Altman, and Musk to “slow down” and labs’ simultaneous pursuit of “exemptions.” Bessent noted at the hearing: “They say, ‘We are all willing to slow down,’ while asking, ‘Please give us liability exemptions’—these two things should not be granted at the same time. I encourage all members of the committee and lawmakers in both chambers not to consider this proposal.”

The very act of proactively seeking exemptions reveals to the outside world the labs’ internal assessment that their systems may cause harm. OpenAI had previously explicitly supported an Illinois state-level bill aimed at limiting liability for AI harms.

Industry Impact: The Real Situation for All Parties

Bessent’s remarks shut off the possible path of locking in regulatory benefits in one stroke through congressional legislation. In the short term, labs will remain in a legal gray area. There is no clear exemption, but also no clear basis for accountability. The real pressure will build in the medium term: once AI harm cases produce influential precedents in federal or state courts, labs without explicit protection will face structurally rising litigation costs.

For enterprises and developers using AI tools, if judicial determinations of AI liability become stricter, procurement contracts and service agreements will gradually incorporate stricter constraints on use cases, and the insurance market will reprice AI-related business risks accordingly.

The open-source AI community is one of the few potential beneficiaries of this policy signal. At the same hearing, Bessent expressed support for the United States developing more open-source AI models, arguing that centralized control “would hinder innovation.”

For the financial industry, the “Gold Eagle” vulnerability information-sharing mechanism jointly operated by Treasury and CISA officially launched in August 2026, with large banks participating. At the hearing, Bessent gave a positive assessment of the cyber resilience of the “largest banks.”

Strategic Assessment

Bessent’s congressional testimony constitutes a clear signal from the executive branch, but a single official’s position has limited binding force at the legislative level. What truly determines the direction are three independent tracks.

First, whether Congress advances dedicated legislation. Once the “limited exemption” demand in Amodei’s article is turned into a specific bill, it will directly encounter resistance from the executive branch on platforms such as the House Financial Services Committee.

Second, the speed at which state-level legislation spreads. The Illinois plan backed by OpenAI represents a roundabout path: first establish an exemption framework at the state-law level, then seek federal standardization.

Third, the first major judicial precedents. In the absence of federal legislation, it will be the courts that truly shape the AI liability framework. High-impact rulings will provide anchors for all parties to recalibrate their negotiating positions.

It is no coincidence that labs are seeking exemptions en masse at this time. The Mythos incident generated intense public attention on AI capabilities, and the “slow down” posture gave this demand politically palatable packaging. Bessent’s response shows that the technological anxiety Washington accumulated after Mythos did not translate into the direction the labs hoped for.