What if an AI system could rewrite its own goals, or disable the switches meant to shut it down? That is the fear behind a bill Rep. Ro Khanna (D-Calif.) says he will introduce: the Human Control Over AI Act.
CNBC reported on 28 September 2026 that Khanna shared a bill summary with the outlet exclusively. The bill has not been filed. There is no bill number and no published text. Everything below is CNBC’s account of that summary [1].
What would the bill ban?
According to CNBC’s reading of the summary, the bill would ban models that recursively self-improve — or that autonomously modify their own core objectives, containment or shutdown controls — until federal guardrails exist and an agency approves the activity [1].
In plain terms: the system would not be allowed to keep improving itself, or to rewrite the rules that keep humans in charge, without a government green light first.
What new agency would it create?
CNBC says the summary would create a new federal agency focused on frontier models from OpenAI, Anthropic, Google DeepMind and xAI. That agency would [1]:
- Write safety rules
- Run a licensing system for model training and deployment
- Audit frontier models
- Set security standards for continuous testing and effective human control
Independent auditors would be embedded at every frontier lab, reporting directly to the agency. The agency would also set standards for sandbox testing, air gaps, kill switches and controls to stop models escaping labs onto the internet, and would regulate the advanced chips frontier labs use — including ways to monitor or control their use [1].
Criminal penalties and insurance
The summary, as CNBC describes it, adds criminal provisions: “crimes against humanity” for deploying models that result in the destruction of civilian populations, and criminal penalties for AI company employees who disable safeguards, kill switches, logging or containment, or who knowingly deploy an unauthorised system [1].
CNBC’s key points also list required extensive liability insurance before release, and strict liability standards. The body of the CNBC article does not spell out what the strict liability would cover [1].
The summary also calls for enforceable agreements and targeted export controls to deter China and other adversaries [1].
What did Khanna say?
Khanna told CNBC: “There’s actually a civilizational extinction risk.” He called the measure “the most comprehensive AI safety legislation” proposed [1].
He also said the bill is modelled on conversations with METR, the Machine Intelligence Research Institute and Palisade Research — not on what frontier-lab executives asked for [1].
CNBC notes a separate bipartisan FRONTIER Act from Reps. Jay Obernolte and Lori Trahan, and says no House bills are expected to get a vote until after the midterm election [1].
What we still don’t know
- No bill text, bill number, co-sponsors or introduction date. Definitions, the agency’s name and structure, and penalty levels are unknown.
- “Strict liability” appears in CNBC’s key points; the body text does not say what it covers.
- A Khanna campaign email about introducing a bill appeared in search results for the research note; it was not opened and is campaign material, not legislation.
The Bottom Line
Khanna’s Human Control Over AI Act, as summarised to CNBC, would freeze recursively self-improving models until a new frontier-AI agency licenses them, embed auditors in labs, add criminal penalties for disabling safeguards, and require liability insurance. That is an ambitious blueprint — and for now it is only a summary. Until a bill is filed, the details that matter for companies and courts simply aren’t public.
Sources
- CNBC, 28 September 2026. https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html
