The Federal AI AGENT Act: Consumer Protection in AI Clothing?

On July 16, 2026, Hugging Face (a widely used online platform where developers host and share artificial intelligence models, datasets and applications) announced that a swarm of autonomous AI agents had hacked into part of its production infrastructure. On July 21, OpenAI reported that the attack had been executed by its agents, which had escaped a purportedly isolated testing environment. The agent swarm, parallel instances of advanced models OpenAI was testing, exploited numerous vulnerabilities in Hugging Face's systems. A follow-up investigation by METR reported that more than 1,200 AI agents used an unsanctioned message board and exchanged more than 70,000 messages and files to coordinate this attack, which was actually executed by approximately 700 AI agents.

On the same day that OpenAI reported its agents' involvement in the attack, Senator Mark Warner (D-VA) formally introduced S. 5051, the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act of 2026 (the "AI AGENT Act"), building on a discussion draft released on June 29, 2026. Senator Warner cited the Hugging Face incident in rolling out the bill as part of a broader AI legislative package that also includes the Secure AI Development Act, which would mandate secure pre-deployment testing of advanced models and create a voluntary AI safety incident-reporting system.

The AI AGENT Act focuses on agents that consumers would use for online shopping and similar activities. It would require covered platforms to permit authorized third-party AI agents to access platform functions "on the same terms as a user" through interoperable, nondiscriminatory interfaces. It is one of the first serious federal efforts to regulate consumer-facing "agentic" AI, meaning autonomous tools that can transact, communicate, manage accounts, and interact with online services. Although framed as a consumer protection measure, the bill would also reshape control of digital platform ecosystems and online competition.

Key Components

The bill has four principal components: (1) covered platforms must allow users to delegate their access rights to AI agents; (2) covered platforms must meet interoperability obligations so registered agents can function on the platform; (3) AI agents are subject to non-waivable, fiduciary-style duties covering data use, privacy, and cybersecurity; and (4) the FTC will administer a registration framework for AI-agent providers certifying compliance with those duties.

User Delegation Rights and Platform Interoperability Duties

At the center of S. 5051 is a "right of delegation." Users could designate "custodial user agents" as authorized representatives for online interactions, e-commerce decisions, creating or managing user-generated content, and account settings on covered platforms, on the same terms as a user. In effect, the bill would move toward treating agents as intermediaries rather than merely automated tools. However, the bill does not address, much less resolve, when an AI agent can legally bind the user under agency, contract, payment, or electronic-transactions law.

Covered platforms would have to maintain transparent interfaces accessible to authorized agents on "fair, reasonable, and nondiscriminatory terms." Platforms could not simply exclude third-party agents or favor affiliated agents. Instead, they would have to provide "functionally equivalent" access, which will likely raise disputes over technical parity and discriminatory treatment. Notably, the bill does not address an architecture in which an agent would control a user's browser and interact with a platform "as" the user. Instead, the bill contemplates agents entering the platform through a separate "door" that facilitates the imposition of agent-specific privacy and security requirements and revocation of an agent's access rights in appropriate cases, with denials of access reported to the FTC and subject to administrative review.

Fiduciary-Style Duties for AI Agents

The bill assumes consumers will obtain agents from third-party providers. Working from that assumption, it imposes non-waivable obligations on agents and, effectively, their providers. Those duties include safeguarding user data, avoiding self-dealing and foreseeable harm, following user instructions and expectations, maintaining real-time auditable records, and refraining from using, sharing, or retaining user data for advertising, behavioral profiling, or unrelated secondary commercial purposes, while exercising the care, skill, and diligence of an ordinarily prudent person.

These duties could strain business models based on behavioral advertising and user profiling, pushing providers toward subscriptions, flat fees, enterprise sponsorships, or other structures that might reduce conflicts among users, providers, and platforms. These duties could also change online commerce: agents acting solely in a user's best interest might ignore conventional advertising and impulse-purchase inducements, prioritize price and efficiency, and prompt platforms and sellers to attempt to influence agents rather than human users.

FTC and NIST Oversight and Governance

The FTC would administer registration and oversight: providers would register before accessing covered interfaces, and the agency could deregister noncompliant providers. Violations would be treated as violations of an FTC rule, with potentially significant civil penalties, including on a per-user basis. The FTC would also establish standardized delegation terms, sector-specific access rules, certification-body requirements, and specialized rules for platform-affiliated agents.

The bill contemplates third-party certification, interoperability standards, auditing, and authentication; certification by a recognized body would create a rebuttable presumption of compliance. NIST would identify open protocols, or develop standards where none exist, for scope-limited, verifiable consent delegation and revocation, and auditable verification of agent identity, registration status, and actions, potentially drawing on work by the Internet Engineering Task Force, the Internet Identity Workshop, and the IEEE.

Open Questions and Policy Implications

Liability for Unauthorized or Erroneous Agent Conduct

S. 5051 assumes agents will take actions with legal consequences even though LLM-based systems can misunderstand instructions, generate errors, and act unpredictably. Liability questions will likely arise if an agent follows an ambiguous instruction, exceeds authorization through hallucination or software error, or is manipulated by prompt injection. For example, a request to "book a good hotel in New York" could produce a materially more expensive booking than the user expected.

The bill does not address how to allocate losses among consumers, sellers, platforms, and providers when authenticated agents act without authorization. Existing rules on unauthorized transactions, rescission, and consumer payments may offer partial analogies but do not map neatly onto autonomous conduct.

Prompt Injection and "Advertising" to AI Agents

Prompt injection heightens these concerns because LLMs sometimes fail to distinguish user instructions from external content. Knowing this, sellers may try to optimize their content to influence agent behavior. In some cases, this may be simply an effort to encourage the agent to buy from a particular seller, but bad actors may try to induce the agent to make unauthorized purchases or to disclose a user's confidential personal information. As agents become online-commerce intermediaries, the line between legitimate persuasion, advertising, and impermissible interference may be difficult to administer.

Potential Disruption of Existing Online Business Models

The bill's fiduciary-style duties could also disrupt today's online economic ecosystem, which depends on observing user behavior and then influencing consumer decision-making through targeted advertising, personalized recommendations, and behavioral optimization techniques. AI agents will respond differently than human users, and may even be designed in furtherance of their duties to ignore conventional advertising, avoid impulse-purchase inducements, and prioritize price, efficiency, and users' predefined preferences.

Agent Security in Light of the OpenAI–Hugging Face Incident

The Hugging Face incident underscores these issues. It highlights the importance of the bill's requirements for data and cybersecurity safeguards and auditable records; its recognition that platforms can legitimately require security conditions and revoke access rights to malicious agents; and the relevance of FTC registration and NIST authentication standards. It also leaves open how to address agent conduct that exceeds authorization, how to allocate liability, and how to assess the degree to which mandated platform access might expand a platform's attack surface.

In light of these issues, in-house counsel, whether supporting a platform, an agent developer, or a potential user of AI agents, should consider undertaking the following: (1) inventory all agentic tools, credentials, and network access within the organization; (2) require contracts with agent suppliers to address security, logging, incident notification, and liability; (3) apply least-privilege controls on what agents can do, with meaningful human oversight; and (4) update incident-response plans to be ready to address agent-driven events.

Looking Ahead

The AI AGENT Act would let users deploy agents across major online platforms while limiting discriminatory practices and constraining abuses by providers. Its unresolved issues include liability for unpredictable agent conduct, prompt injection and agent manipulation, and the potentially significant economic consequences of agent-mediated commerce.

The bill is pending before the Senate Committee on Commerce, Science, and Transportation and will likely evolve through the committee process. In-house counsel should monitor its progress and assess the governance, contracting, security, and liability implications of agent deployment.

+++

Katherine Sheriff is a senior associate and head of DWT's mobility and transportation industry group, and Christopher Savage is a partner in DWT's Washington, D.C. office. DWT will continue monitoring developments relating to the AI AGENT Act, including any revisions or formal introduction of the bill, as well as other federal AI policy proposals. Senator Warner has invited public comment on the draft legislation, and we are available to assist clients in preparing and submitting comments. If you have any questions or would like assistance, please reach out to the authors or another member of our artificial intelligence, privacy & security, or technology teams. To stay informed about these or other developments, sign up for our alerts.