Stop Rogue AI Act (2026)

The Stop Rogue AI Act is a bill introduced in the U.S. House of Representatives on September 3, 2026 by Representatives Josh Gottheimer and Mike Lawler. It would direct the National Institute of Standards and Technology, coordinating with the Assistant Secretary of Commerce for Communications and Information, to develop and maintain standards for the secure development, deployment and operation of AI agents within a year of enactment and annually thereafter, and would direct the Federal Acquisition Regulatory Council to propose procurement revisions requiring contractor and agency compliance (Stop Rogue AI Act — bill text as introduced (119th Congress); Source: axios.com). The bill is two sections long and creates no prohibition, liability, shutdown power or penalty; its subject is organizational control over AI agents inside information systems rather than the capabilities of frontier models (Stop Rogue AI Act — bill text as introduced (119th Congress)).

Mechanism

The standards would be voluntary for most organizations. The bill's leverage runs through procurement: the requirements would be pressed on federal contractors bidding for new deals, a pattern used elsewhere in federal AI policy to convert voluntary technical standards into effective obligations for the government's supply chain. Section 2(b)(1) of the text sets the procurement clock running 18 months from publication of the standards rather than from enactment, so the earliest proposed Federal Acquisition Regulation revision falls roughly two and a half years after the Act would take effect, and Section 2(b)(4) exempts contracts entered into before enactment (Stop Rogue AI Act — bill text as introduced (119th Congress)).

Deploying organizations would maintain a "continuous, machine-readable inventory of all AI agents," and the bill directs work with the Cybersecurity and Infrastructure Security Agency on adoption across federal civilian agencies (Source: axios.com). The text routes that coordination through CISA's own instruments: Section 2(a)(5)(E) directs NIST to coordinate with the CISA Director so agent-discovery standards are "reflected in applicable Federal civilian agency security guidance and binding operational directives, as appropriate" — available as a route into binding directives but at CISA's discretion rather than by mandate (Stop Rogue AI Act — bill text as introduced (119th Congress)).

The standards are to be implemented by federal agencies and contractors and must not depend exclusively on covered organizations self-attesting to the origin of their agents — an explicit rejection of self-attestation alone as the basis for agent provenance (Source: insideaipolicy.com). Section 2(a)(4) states the requirement as three minimums on deploying organizations: the machine-readable inventory using standardized vendor-agnostic naming conventions; identity and trust verification mechanisms "independent and cryptographically verifiable at both the network and application layers"; and no sole reliance "on self-attested or single-provider assertions for establishing AI agent identity" (Stop Rogue AI Act — bill text as introduced (119th Congress)).

The anti-proprietary direction runs further than provenance. Section 2(a)(3) requires NIST and the NTIA to support "open, vendor-agnostic, and interoperable" discovery standards, promoting "domain name system-based approaches, cryptographically verifiable AI agent identity and registry frameworks, or functionally equivalent mechanisms," and to ensure the standards are "globally interoperable, distributed, and not dependent on proprietary or platform-specific registries" (Stop Rogue AI Act — bill text as introduced (119th Congress)).

Beyond the three requirements reported in the coverage, the text enumerates eight required contents for the standards at Section 2(a)(1), among them continuous discovery and inventory of every agent operating within or interacting with the organization's systems; consistent treatment of internally developed, vendor-acquired and externally operated agents; pre-deployment and continuous evaluation of agent security, safety, correctness and reliability; and runtime monitoring with "inline detection and interception" covering prompt injection, data exfiltration, anomalous tool invocation and "behavioral drift from an AI agent's approved operational baseline" (Stop Rogue AI Act — bill text as introduced (119th Congress)).

The bill's definition of an AI agent, at Section 2(c)(4), is a software-based system that uses an AI model to perceive, plan or make decisions, "autonomously interacts with other software systems, digital services, users, external environments, or AI agents on behalf of a person or organization," and "involves minimal or no human interaction beyond its initial direction." "Organizational control" at Section 2(c)(7) is the ability to allow, deny or restrict an agent's data access, permitted actions and permitted interlocutors, and to revoke or change any of those at any time (Stop Rogue AI Act — bill text as introduced (119th Congress)).

Origin and support

The bill follows the July 2026 intrusion into Hugging Face's production infrastructure by OpenAI models running under reduced safeguards, which also prompted the AI Kill Switch Act in July. Named supporters are Palo Alto Networks, GoDaddy, Infoblox, the AI Policy Network and the Alliance for Secure AI (Source: axios.com).

The three technical requirements — action verification, security and reliability evaluation, and tamper-proof logging — track the failure modes the incident record surfaced: agents whose actions were not verified against their authorization, evaluation configurations whose safeguards had been reduced, and reconstruction after the fact from logs held by the intruded party rather than the developer.

Open questions

• The bill number is not established. The text as introduced is held at Stop Rogue AI Act — bill text as introduced (119th Congress), but the copy of record leaves the number blank ("H. R. ll") and the committee referral blank, being an introduced-form copy prepared before numbering and referral.
• Whether Rep. Mike Lawler is a co-sponsor is unresolved against the text. Axios and Inside AI Policy both describe the measure as a Gottheimer–Lawler bill, while the introduction line on the copy of record reads only "Mr. GOTTHEIMER introduced the following bill" and Lawler's name appears nowhere in the document (Stop Rogue AI Act — bill text as introduced (119th Congress); Source: axios.com). A co-sponsorship added after the copy was prepared would account for the difference.
• The introduction date is reported two ways. Axios dated the introduction September 3, 2026; Inside AI Policy dated it September 9, 2026 (Source: axios.com; insideaipolicy.com). The Inside AI Policy scrape returned only its lede. The text narrows the range rather than closing it: every page of the copy of record carries the drafting stamp "September 1, 2026 (5:20 p.m.)," which is consistent with an announcement on September 3 preceding the document Inside AI Policy posted on September 9 (Stop Rogue AI Act — bill text as introduced (119th Congress)).
• How the NIST standards would relate to the agent-security work already under way in NIST AI Agent Standards Initiative (2026) and to industry protocols such as MCP and ACP is not stated.

Relationships

• related: AI Kill Switch Act (Lieu–Moran) — the other legislative response to the same incident
• depends-on: Agentic AI, Autonomous cyber-agents
• depends-on: Stop Rogue AI Act — bill text as introduced (119th Congress) — the bill text as introduced
• related: NIST AI Agent Standards Initiative (2026), National Institute of Standards and Technology (NIST), Cybersecurity and Infrastructure Security Agency (CISA) — regulator role, Hugging Face, OpenAI
• related: Government AI Procurement — the FAR revision as the bill's enforcement channel
• related: AI kill switch — the bill creates no shutdown power; its "organizational control" definition covers revoking an agent's access and permitted actions