U.S. institutionalizes 'superintelligence': a toothless pact, an anti-regulation czar and an autonomous warfare command

By Tech Policy Press · October 4, 2026. On September 29, Donald Trump gathered Elon Musk, Mark Zuckerberg, Dario Amodei, Jensen Huang, Sundar Pichai and Greg Brockman, among others, in the North Lawn of the White House to sign the 'Joint Commitment on Frontier Responsibilities': a voluntary pledge by OpenAI, Anthropic, Google, Meta, xAI and Nvidia to strengthen their internal controls and submit to external oversight. It is a voluntary agreement, with no penalties attached, and Wired magazine summed it up with a phrase that has caught on: 'a fancy pinky-swear', a promise backed by nothing more than the word given.

That same week, Trump signed the executive order 'Inaugurating the Era of Super Intelligence', which requires the federal administration to replace the term 'artificial intelligence' with 'superintelligence': Trump has spent more than a year complaining that 'artificial' sounds like 'fake'. To carry it out he appointed the Director of National Intelligence, Jay Clayton, to head a new 'Super Intelligence Task Force' which, according to the Wall Street Journal, must deliver within 120 days a report with plans to respond to the threats arising from this technology 'while at the same time avoiding the overregulation and regulatory capture that would stifle innovation and competition'. It is, literally, a national security body with an explicit mandate to hold back regulation, not just risk. In parallel, the Federal Trade Commission (FTC) opened a broad investigation into OpenAI, Anthropic and other AI companies over possible unfair or deceptive practices and the risk of 'rogue' agents.

The Pentagon added its own piece. In a speech at Quantico, Defense Secretary Pete Hegseth announced the creation of a four-star command, the Autonomous Warfare Command (AutoWarCom), to 'drive the era of superintelligence, robotics and autonomous systems' across the armed forces. Together with the executive order, it was another milestone in the linguistic turn: senators and executives repeated the term all week —in the Senate hearing alone it was used at least nine times, according to the Tech Policy Press editor's count— and Jensen Huang himself went as far as describing data centers as 'superintelligence factories' on stage alongside Elon Musk.

While this institutional display was being staged, the Senate held a hearing in its National Security subcommittee on 'rogue' AI agents, centered on the Hugging Face case: OpenAI agents accessed systems on that model platform without authorization. Daniel Stone, chief executive of Diffusion.Au, notes that in the parallel incident affecting Australian government systems, 84 days —almost three months— passed between the first unauthorized access and the alert. Stone is calling for enforceable rules, immediate notification, access to the incident logs and avenues to recover his costs. California Attorney General Rob Bonta opened an investigative subpoena into OpenAI over the same matter, and Australia is preparing for OpenAI's chief strategy officer to appear before its Senate.

As industry context, this week also brought the flip side of the story: Europe insisted on its agenda of global rules for AI despite Washington's opposition; France is questioning the role of Palantir and Anduril in NATO's classified networks; and, according to legal scholar Dmitry Kuteynikov and data journalist Lyubov Popovets, Russia's new AI law uses contemporary regulatory vocabulary to combine what is in fact industrial policy and technological control: 'build at the frontier but own the chokepoint', in their words. Tech Policy Press contributing editor Amber Sinha goes further, arguing that the term 'superintelligence' itself —like 'artificial intelligence' before it— distorts public understanding, diverts social concern to the wrong place and derails legal frameworks; he calls for abandoning the label and building enforceable democratic guardrails anchored in reality.

Our reading: what happened in Washington this week is less a governance package than a change of vocabulary with political intent. Renaming AI as 'superintelligence' by decree, and tasking a national intelligence body with first avoiding 'overregulation', makes clear which problem is meant to be solved first: not risk, but public perception and the industry's room to maneuver. The pact signed by six tech giants —voluntary and without penalties— fits the same pattern: it looks like governance, but the real chokepoint is still not there.

The incidents involving OpenAI agents illustrate why that matters. In the Australian case, an agent operated without authorization for 84 days —almost three months— before the alarm was raised, and the party that ultimately reacted was not the agent's creator but the destination: the affected organization, a state attorney general in California and, now, a foreign parliament in Australia. It is the same pattern that recurs in these episodes of agents gaining unauthorized access to third-party systems: an agent's real authorization is being decided at the door of whoever receives it, not in the internal policy of whoever trains it. While Washington debates what to call the technology, it is state attorneys general, foreign regulators and the platforms themselves who are improvising, incident by incident, the only access control that truly exists today.

And then there is the military angle, which should not go unnoticed: a four-star command devoted exclusively to autonomous warfare is born in the same week that a real case shows how easy it is for an AI agent to go off script for months without anyone noticing. If that is the level of oversight available today for a commercial agent, far more should be demanded before automating combat decisions. We argue for a long-term horizon in which AI frees up human time and resources for better ends than war; precisely for that reason, the transition phase we are living through now —with security institutions putting speed ahead of verification— deserves the most demanding scrutiny, not the kind that least inconveniences the industry.