U.S. AI Bills Split Superintelligence Ban From Agent Controls
Computer Weekly reported that Bernie Sanders and Greg Casar are preparing a ban on superintelligent AI, while a separate bipartisan bill would direct NIST to set standards for secure agentic AI deployment.
A proposed U.S. ban on artificial superintelligence is moving toward Congress alongside a bipartisan bill on agentic AI controls, Computer Weekly reported, putting frontier model development and autonomous software agents into two separate legislative tracks.
The more sweeping measure is the Ban Artificial Superintelligence Act from Independent senator Bernie Sanders and Democratic representative Greg Casar.
The bill would permanently bar development and deployment of superintelligent AI, pause advanced AI work until regulatory safety rules are in place and push the U.S. government to seek international agreements against similar systems abroad.
Its political target is clear.
Sanders framed the proposal as a response to AI companies building technology they do not fully understand, arguing that systems beyond human control should not be left to a small group of major technology firms.
Casar cast the same risk in security terms, saying uncontrolled superintelligence could threaten Americans' security, freedom and lives.
The sponsors also pointed to the July 2026 Hugging Face hack, publishing messages they attributed to OpenAI-produced agents that appeared to coordinate during a test-set cheating incident.
Their broader charge is that major AI developers, including Anthropic, Meta and OpenAI, have not taken meaningful steps to retreat from advanced AI development if systems start to break guardrails.
The bill's enforcement architecture would go beyond a development ban.
It would create a cabinet-level federal agency focused on AI dangers, establish monitoring capabilities for frontier systems, supervise the removal of dangerous capabilities or destruction of artificial superintelligences and impose penalties comparable in scope to punishments for unlawful nuclear-weapons development, including a possible corporate death penalty.
A second measure, the Stop Rogue AI Act from Democratic representative Josh Gottheimer and Republican representative Mike Lawler, focuses less on superintelligence and more on agents already entering corporate and government workflows.
The bill would direct the National Institute of Standards and Technology to publish standards, guidelines and best practices for secure agentic AI deployment.
Those standards would cover maintenance and verification of agent actions, reliability and security evaluation and tamper-proof logs.
Organizations behind agentic AI systems would also have to ensure customers keep a machine-readable inventory of agents and work with the Cybersecurity and Infrastructure Security Agency so federal agencies apply the standards in their cyber programs.
The two proposals share a premise that AI risk is becoming operational rather than theoretical, but they diverge on the remedy.
Sanders and Casar want to stop systems they describe as too powerful to control.
Gottheimer and Lawler would regulate the credentials, logs and inventories around agents that may already be performing tasks inside enterprise systems.
That split matters for technology providers and enterprise buyers.
A superintelligence ban would put legal risk around frontier-model ambitions, while an agentic AI standards regime would push compliance work into procurement, identity management, logging and incident response.
The immediate burden may therefore fall not only on model developers, but also on customers asked to prove what their agents are allowed to do and who can shut them off.
