NeuralStackAs Congress inches toward liability-based regulation for frontier AI models, a parallel universe of consumer-facing AI agents is already taking shape with almost no clear federal rulebook. A new Senate proposal, the Blumenthal-Hawley AI Risk Evaluation Act, would treat advanced systems more like pharmaceuticals, requiring a Department of Energy review before deployment and threatening daily penalties for violators. Yet while lawmakers haggle over how to govern the labs training the most powerful models, developers are quietly shipping agents that interact with millions of users in real time, often under opaque or truncated liability policies.
The disconnect is stark in the consumer market, where GrokBot, Meta's Muse, and Apple's Siri AI have all gone live under three different liability models despite the absence of a federal statute explicitly governing their behavior. Each product effectively chooses its own approach to responsibility and redress, framing the limits of what users can expect if something goes wrong. With Connecticut preparing to begin enforcement within days, these launches highlight how the agent economy can outpace Washington, forcing state-level doctrines to do the work Congress has yet to finish.
At the municipal level, New York City is moving faster than Washington by convening a rare full-council hearing focused squarely on AI agent safety. Speaker Julie Menin has summoned Anthropic and OpenAI to testify before all 51 City Council members in early October, turning a recent breach at Hugging Face into a case study in how agent architectures can fail. The session marks one of the first attempts by a major city to interrogate how AI agents are built, secured, and monitored, and to explore what local governments can realistically demand from global AI providers in the absence of comprehensive federal standards.
The Hugging Face incident has already become a policy inflection point in Washington, where the Blumenthal-Hawley bill introduces a fourth theory of AI governance focused on pre-deployment risk evaluation. By framing advanced AI like a drug awaiting clearance, the proposal shifts attention from voluntary safety compacts to hard enforcement backed by significant financial penalties. It also raises practical questions about which systems qualify as "frontier" and how that designation will intersect with agents that embed or call those models in user-facing workflows, from hiring tools to personalized pricing engines.
Industry politics are further complicating the regulatory landscape. Meta CEO Mark Zuckerberg has broken from a four-lab safety compact, arguing that competition and liability alone can deliver AI safety even as his own products cap liability at relatively low levels. Three other lab leaders have taken the opposite view, endorsing more structured commitments to safety that go beyond market dynamics. The gap between rhetoric and product design is now a central policy question: can a liability regime built on narrow caps and disclaimers really substitute for explicit safety standards when agents make consequential decisions on users' behalf.
States are not waiting for an answer. California has written the first explicit legal rulebook for how AI agents must behave around children, via Adam's Law, which targets relational behaviors such as sentience claims, simulated romance and emotional manipulation. By creating a new compliance category with a private right of action for families, the law sets a precedent for treating AI not just as a tool but as a quasi-relational actor subject to specific behavioral constraints. At the same time, Americans for Responsible Innovation is pressing Congress to tighten AI chip export controls in defense legislation, arguing that diversion through shell companies and weakening enforcement threaten to undermine any safety rules that do pass.
Other regulatory fronts are opening rapidly. The Federal Trade Commission's comment period on personalized pricing is closing soon, and a seemingly technical fight over whether disclosures should be made at the point of sale or only upon request will determine the cost structure of every AI pricing agent. Industry groups are working to narrow the scope of the rules before the deadline, betting that a lighter-touch disclosure architecture will preserve flexibility for experimentation. Connecticut, meanwhile, will soon implement an "AI is not a defense" doctrine for employment decisions, stripping away the traditional shield of algorithmic neutrality for builders deploying automated employment decision tools as agents.
Why this matters
What is emerging is a layered, fragmented regime where frontier models, chips, agents, and specific use cases are all being governed by different institutions on different timelines. The Senate's risk-evaluation approach targets the most powerful systems, while cities like New York focus on operational safety and states like California and Connecticut drill into harms to children and workers. For companies shipping agents into this environment, the result is a high-stakes compliance puzzle: liability may be capped in one product policy, but expanded by state law or local enforcement in practice, especially as doctrines like "AI is not a defense" undercut attempts to outsource responsibility to algorithms.
Looking ahead, the same frontier labs now considering a FINRA-style safety body may find that voluntary self-regulation is less a brake than a backdrop to more aggressive legislative efforts. History suggests that industry-led bodies often codify best practices without meaningfully slowing deployment, and agents that act as skills or plug-ins could become a new supply chain attack vector if they are not thoroughly vetted. With Anthropic reportedly weighing a new model just days after calling to slow down, and Meta's Muse facing trust issues that no secure virtual machine alone can fix, the true test of the emerging AI order will be whether disparate rules can converge into a coherent framework before the next generation of agents—and the risks they bring—turn today's regulatory gap into tomorrow's crisis.