Khanna Bill Bans Recursive AI Until Federal Guardrails
Silicon Valley Democratic Representative Ro Khanna will introduce the Human Control Over AI Act, a bill that would ban recursively self-improving AI and autonomous changes to a model’s.
PromptCrates Editorial
Staff Writer
Silicon Valley Democratic Representative Ro Khanna will introduce the Human Control Over AI Act, a bill that would ban recursively self-improving AI and autonomous changes to a model’s core objectives, containment, or shutdown controls until federal guardrails exist and an agency approves those activities, CNBC reported exclusively on 28 September 2026. Khanna told CNBC there is “civilizational extinction risk,” plus separate safety risks of loss of control and misuse, and that both must be taken seriously.
What the bill would ban and create
The summary shared with CNBC centers on recursive self-improvement. Until a new federal agency writes rules and licenses the work, companies could not deploy models that recursively self-improve or that autonomously modify their own core objectives, containment systems, or shutdown controls. That targets the same automated R&D feedback loop researchers warned about the same day in the intelligence-explosion paper covered on PromptCrates as AI leaders on automating R&D.
The bill would create a federal agency focused on frontier labs including OpenAI, Anthropic, Google DeepMind, and xAI. Duties would include safety regulations, a licensing system for model training and deployment, frontier model audits, and security standards for continuous testing and effective human control. Independent auditors would be embedded at every frontier lab and report directly to the agency. Standards would cover sandbox testing, air gaps that isolate test environments from the internet, kill switches, and controls meant to stop models from escaping labs onto the public internet. The agency would also regulate advanced chips used by frontier labs, including monitoring or control mechanisms.
Khanna called the package “the most comprehensive AI safety legislation” proposed and said it is modeled on conversations with safety nonprofits such as Model Evaluation and Threat Research (METR), the Machine Intelligence Research Institute, and Palisade Research—not on wish lists from frontier-lab executives. “I want this to be a model to give voice to the AI safety community,” he told CNBC, arguing that community has been unfairly dismissed as science fiction.
Liability insurance and criminal penalties
Beyond licensing, the bill hardens legal consequences. AI companies would need extensive liability insurance before releasing models. It would criminalize “crimes against humanity” for deploying AI models that result in the destruction of civilian populations. It would also impose criminal penalties on employees who disable safeguards, kill switches, logging, or containment systems, or who knowingly deploy an unauthorized system. Those clauses aim at insider circumvention as much as corporate negligence.
Foreign-policy tools appear as well. The bill calls for the administration to pursue enforceable agreements and targeted export controls to deter China and other adversaries from developing dangerous AI systems. That pairs domestic licensing with outbound chokepoints on chips and know-how—an approach familiar from existing semiconductor export policy, now extended explicitly to frontier AI risk.
The proposal joins other congressional efforts, notably the bipartisan FRONTIER Act from Representatives Jay Obernolte, R-Calif., and Lori Trahan, D-Mass., which would place independent auditors in frontier labs and allow the government to shut down models with catastrophic-risk potential. PromptCrates readers comparing standards-body paths can also revisit SAFA frontier AI standards authority coverage and the NIST agent standards gap. CNBC notes that no House bills are expected to receive a vote until after the midterm election, and the Senate is expected to leave Washington after this week until after the election—so the near-term impact is agenda-setting, not immediate law.
Political timing and industry reaction paths
Khanna’s introduction lands in a week when OpenAI publicly held GPT-6.1 Astra for safety shortfalls and senior researchers warned about intelligence explosions. That news cycle strengthens the bill’s narrative that voluntary lab caution is not enough without statutory authority, licensing, and embedded auditors. Labs that already accept independent evaluators may argue they are halfway there; critics will ask whether a new agency and recursive-AI ban would freeze beneficial research or mainly entrench compliance-heavy incumbents.
For enterprise counsel, the watch items are definitions. How “recursive self-improvement” and “autonomous modification of containment” are written will determine whether ordinary fine-tuning, online learning, or agent self-prompts fall inside the ban. Chip-monitoring authority could also reach cloud customers who rent frontier training clusters. Until text is introduced and marked up, treat CNBC’s summary as directional: ban-until-licensed recursive AI, new agency, embedded auditors, insurance mandates, and severe criminal exposure for disabling safeguards.
Primary reporting for this article: CNBC’s 28 September exclusive by Garrett Downs on Khanna’s Human Control Over AI Act summary and interview. Facts stay anchored there: ban on recursive self-improve and autonomous objective/containment/shutdown changes until guardrails and approval; new agency for OpenAI, Anthropic, Google DeepMind, and xAI; licensing, audits, sandbox/air-gap/kill-switch standards; chip controls; embedded auditors; liability insurance; crimes-against-humanity and employee criminal penalties; China-focused agreements and export controls; METR/MIRI/Palisade inspiration; FRONTIER Act context; post-midterm vote timing.
