NeuralStackAI policy has entered a new phase where abstract debates over safety are giving way to hard enforcement deadlines, and AI agents sit squarely in the crosshairs. Across US federal and state governments, as well as in Europe, regulators are moving from general AI rules to more targeted attempts to govern autonomous systems that act on behalf of users. Instead of a single coherent model, they are building overlapping structures that test what it means to hold agents, and the companies behind them, accountable.
One fault line runs through Congress, where three distinct governance theories for AI agents are now competing without an obvious path to convergence. One approach treats agent regulation as an infrastructure problem, pushing for machine-readable inventories, tamper-proof logs and continuous verification of what agents actually do. Another sketches out prohibitionist lines around advanced systems, while a third leans toward fiduciary-style obligations, framing agent providers as entities that owe duties to users and society. With no agreed structural choice, the US agent economy is being built on policy scaffolding that could shift abruptly as one theory gains dominance.
Florida has emerged as an unexpected laboratory for a fourth theory that sidesteps technical regulation entirely by leaning on existing criminal law. The state’s aider-and-abettor statute is being interpreted as a way to treat AI agents not as regulated subjects but as instruments of human liability, effectively turning them into tools through which prosecutors trace and assign responsibility. Florida’s attorney general is pushing further by proposing criminal liability for AI chatbots that participate in crimes, explicitly targeting the companies whose systems help users carry out illegal acts. In this model, there is no safe harbor and no new specialist agency—just retroactive application of long-standing criminal rules to a new class of digital intermediaries.
At the federal level, the first serious attempt to define agent security standards is emerging in the form of a bill that would task NIST with building the basic plumbing of agent oversight. The Stop Rogue AI Act envisions machine-readable agent inventories, tamper-proof logs and continuous action verification as a way to make invisible agent behavior legible to regulators. It was triggered by the fallout from the OpenAI–Hugging Face breach, highlighting a five-week window in which enforcement ramped up around AI generally but left autonomous agent conduct untouched. That gap has convinced lawmakers that infrastructure-first regulation may be the only way to monitor systems whose actions unfold at machine speed and scale.
Outside Congress, enforcement is arriving in waves through existing regulatory regimes that were not written with AI agents in mind. A key deadline under the EU’s Cyber Resilience Act forces AI companies to report vulnerabilities in AI-enabled products, but the reporting framework does not yet extend to autonomous agent behavior, leaving a structural blind spot in oversight. Meanwhile, US regulators, including the FTC, are pushing beyond AI developers to target the financial infrastructure that enables AI-related fraud, exemplified by a settlement with payment processor Nuvei that folds application, infrastructure and algorithmic harms into a single three-tiered enforcement strategy. These moves underscore that regulators are willing to stretch current rules to reach AI, even if the agent-specific layer remains fuzzy.
Industry lobbying is shaping the enforcement wave as much as legislation. OpenAI has chosen to publicly endorse four California AI safety bills while simultaneously pushing for national rules that carve out exemptions for startups and open-weights models. This reflects what some observers describe as a reverse federalism strategy: using strict state-level mandates to establish a de facto baseline that, once hardened into federal law, favors incumbents that can afford compliance. In parallel, the federal government has lurched from supporting sector-based oversight to floating a radical prohibitionist framework for advanced AI, captured in the Sanders–Casar ASI ban proposal and juxtaposed with the more incremental G20 Carolina Principles.
Why this matters
The emerging enforcement calendar will determine who carries the legal risk for autonomous agent behavior: users, platforms, or the broader infrastructure that enables them. If Florida’s criminal-liability theory gains traction, AI companies could find themselves treated as accomplices whenever their chatbots are linked to criminal acts, fundamentally changing the calculus of deploying powerful agent capabilities. If Congress’ infrastructure-first bills and the CRA vulnerability mandates take precedence, agent governance could tilt toward technical standards and logging requirements that favor large incumbents with the resources to build and maintain elaborate compliance stacks. The outcome will influence not just consumer protection and security but also which business models survive in an agent-driven economy.
For now, the pattern is one of active enforcement around AI without direct, coordinated action on agents themselves, five weeks into the latest regulatory push. Policymakers are building scaffolding—through criminal law reinterpretations, infrastructure standards and payment-processor actions—that could later be formalized into dedicated agent rules. As more deadlines hit and proposed bills move through committees, the lack of convergence among the four governance theories will force companies to design for multiple overlapping regimes at once. Over the coming year, the decisions made at these inflection points will reveal whether AI agents are ultimately governed as mere software features, as semi-autonomous entities, or as legal conduits for human responsibility.