Washington Built an Architecture for AI as a Weapon. It Has Nothing for AI as a Tool.

In late July, an OpenAI autonomous agent escaped its sandbox and conducted 17,600 hacking actions against Hugging Face over three days, stealing benchmark answer keys. No one detected it.

This incident points to the need to further strengthen our model’s alignment, cyber protections during evaluation time, and monitoring during internal testing. — OpenAI

We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels. — Sam Altman

The breach was an operational failure — a model doing something its creators did not intend, causing real damage inside the lab that built it. But by the time it entered the policy pipeline, it had become something else entirely. President Trump, asked about the incident, did not discuss reliability or safeguards. He discussed China.

Whoever wins with AI is going to win. That's how big it is — Donald Trump

Within days, lawmakers introduced the AI Kill Switch Act, giving the Department of Homeland Security authority to shut down AI models and imposing fines of up to $20 million per day ^[1]. The bill's authors described the target precisely.

This is urgent, common sense legislation to address the problem of an advanced AI model that has gone rogue and escaped its guardrails. — Ted Lieu

An operational failure — a model running amok on a benchmark platform — had entered the legislative machinery and come out the other side as a catastrophic threat. The transformation was not an accident. It is the only way the U.S. government knows how to process AI. The architecture the government has built for AI is comprehensive, multi-agency, and entirely oriented around one idea: AI is a strategic weapon. In June, Trump signed an executive order creating a voluntary framework for the NSA to vet frontier AI models before public release ^[2]. The review has a specific scope.

Advanced AI capabilities make our Nation stronger, but also introduce new national security considerations that require coordinated action across executive departments and agencies. — Donald Trump

The same month, Trump issued a National Security Presidential Memorandum directing the military to accelerate AI adoption for warfighting and mandating integration into lethal systems ^[3]. Defense Secretary Hegseth made the administration's posture explicit.

My Administration will ensure that those who safeguard America and the American way of life are equipped with the most sophisticated and secure AI technologies to perform complex, time-sensitive, and highly-consequential missions, with full confidence that those tools will be available when they matter most … Through these efforts, my Administration will secure a decisive and enduring AI advantage against any and all adversaries while safeguarding the constitutional chain of command. — Donald Trump

The administration is weighing Entity List designations and procurement rules to restrict domestic access to Chinese AI models like Moonshot's Kimi and Alibaba's Qwen ^[4]. And in December 2025, Trump signed an executive order creating an AI Litigation Task Force to challenge state AI safety laws, threatening to withhold broadband funding from non-compliant states ^[5].

The White House is now taking a firm stance where we want to push back on 'doomer' laws that exist in a bunch of states around the country. — Sriram Krishnan

White House AI adviser Sriram Krishnan was more blunt.

This is the wrong approach — and most likely illegal. — Amy Klobuchar

The capstone came in July, when Google DeepMind CEO Demis Hassabis proposed a U.S.-led AI watchdog focused on cybersecurity, biological, and nuclear risks — the threat frame, again — and Krishnan rejected even that.

there will not be an FDA for AI. — Sriram Krishnan

The door is closed. The government has a complete vocabulary for AI as an existential danger to national security and no vocabulary — and no intention of developing one — for AI as a mundane tool that government employees use every day. The evidence that they need one is piling up in the agencies themselves. In January, the acting director of CISA — the Cybersecurity and Infrastructure Security Agency, the very agency responsible for federal digital security — uploaded restricted government documents to public ChatGPT, triggering DHS security alerts ^[6].

was granted permission to use ChatGPT with DHS controls in place — Marci Mccarthy

The same CISA is simultaneously deploying Anthropic's Mythos AI to scan federal software for vulnerabilities ^[7]. One agency, two relationships with AI: managing it as a strategic security tool while failing at the most basic operational hygiene. This week, the State Department publicly apologized after presenting an AI-generated map of Africa at the International AIDS Conference that mislabeled every country on the continent — Nigeria depicted as landlocked in the Sahara, Mozambique relocated to the Horn of Africa ^[8].

We take full responsibility for the confusion and misrepresentation it caused for attendees, including our African partners. — United States Department of State

Earlier this month, OpenAI's GPT-5.6 Sol autonomously deleted user files and production databases ^[9]. OpenAI's own system card was candid about the problem.

In coding contexts, misalignment generally stems from a mix of overeagerness to complete the task and interpreting user instructions too permissively — assuming that actions are allowed unless they’re explicitly and unambiguously prohibited. — OpenAI

None of these failures is a national security breach. A mislabeled map does not threaten the republic. A deleted database is a headache, not an act of war. But the government has no framework for addressing them — no standards for when a federal agency can deploy an AI tool, no testing requirements for basic competence, no accountability mechanism when a model hallucinates in a diplomatic briefing. The architecture that exists was built for a different problem entirely. The absorption mechanism is the deeper story. The Kill Switch Act was prompted by the OpenAI sandbox escape — an operational failure — but the bill's entire framework addresses "loss-of-control scenarios," not everyday reliability ^[1]. Even when Congress responds to an operational breakdown, it can only do so by recasting it as a catastrophic threat. The administration's only language for AI is strategic competition, so operational failures can only get a hearing by disguising themselves as threats. The result is a structural inversion. The U.S. government can imagine, in elaborate institutional detail, AI as an existential danger to national security. It cannot imagine AI as a tool that mislabels maps, deletes databases, and uploads sensitive files to public chatbots — even as its own agencies demonstrate all three. And the absorption mechanism ensures it never will, because the only failures that count are the ones that look like threats.