There is a peculiar and troubling rhythm to governance in the spring of 2026: the machinery of state grinds forward on matters both trivial and vital, often unable to distinguish between them. This week, Congress juggled King Charles's address to the joint session, debated ethanol subsidies in the Farm Bill, prepared votes on FISA reauthorization and military action in Cuba, and renamed the Department of Defense without apparent irony. Meanwhile, in classified briefings that few members will ever attend, executives from OpenAI and Anthropic showed congressional staff the architectural blueprints for potential catastrophe—and then left the room, leaving lawmakers to contemplate the implications alone.
The contrast between the visible and the hidden work of government could hardly be starker. What emerged from Thursday's briefings, pieced together from a cautious Axios report, painted a picture of technological capability racing so far ahead of regulatory capacity that the gap itself has become the central security threat. Anthropic has withheld its Mythos Preview model from public release entirely, citing the speed at which it can identify and exploit critical infrastructure vulnerabilities. OpenAI has adopted a tiered release strategy for its GPT-5.4-Cyber model, granting federal agencies preview access while the public remains in the dark. Both companies frame this restraint as responsible stewardship. The reality is grimmer: private corporations have unilaterally assumed the role of gatekeepers, deciding which government agencies deserve access to which tools and at what pace.
This is not partnership between industry and government, whatever House Homeland Security Chair Andrew Garbarino might have said in his statement. Partnership implies a meeting of equals. What we have instead is a tacit acknowledgment by both sides that government has already lost the race to understand, let alone regulate, the technology reshaping the nature of risk itself.
The Briefing Nobody Understands
Rep. August Pfluger emerged from an earlier briefing on jailbroken AI models visibly shaken. "What I just saw in there, with just a short amount of time typing in questions, is very scary," he told reporters. That a member of Congress tasked with overseeing national security admits that a few minutes with an AI tool terrifies him is perhaps the truest statement made about artificial intelligence policy this year. It speaks to the vertigo induced by knowing that a dangerous thing exists, can cause immense harm, and that no one in the room has any real idea how to stop it.
Rep. Andy Ogles followed with a darker observation: the most dangerous AI tools are already off-the-shelf and easy to access. "It increases the probability that the wrong person gets it," he said. This is the central problem that no classified briefing can solve and no tiered release strategy can prevent. You cannot un-invent a capability. You cannot prevent bad actors from obtaining tools once they exist in the world. The best that private companies can offer is delay, sequencing, and hope that in the time gained, someone figures out what to do.
Yet the briefings themselves reveal how unequipped Congress truly is to grapple with this problem. The very classification of the briefings—the necessity of removing them from public view, the sequestering of information—mirrors a kind of security theater that dates to the Cold War. We protect the information from our own citizens as if doing so prevents foreign adversaries from obtaining it. Meanwhile, China allegedly operates at "industrial-scale" campaigns to distill and copy American AI models, according to a White House memo discussed during the briefings. The classification does nothing except ensure that the American public cannot participate in democratic deliberation about the terms on which AI development should proceed.
The Backhanded Sympathy of Private Constraint
There is something quietly audacious about the position Anthropic and OpenAI have adopted. By declining to release certain models or by controlling access through a tiered system, they position themselves as responsible stewards of technology too dangerous for immediate public distribution. They are, in effect, making government decisions without government accountability. And because Congress lacks the expertise to contradict them, because the executive branch must negotiate rather than command, and because the public has no seat at the table, these companies have become de facto regulators of their own industry.
This arrangement suits everyone except democracy. It suits the companies, which avoid the blunt instrument of legislative restriction and instead maintain the appearance of cooperation while controlling the pace of revelation. It suits Pentagon officials and intelligence agencies, who gain privileged access to dangerous tools before their competitors do. It even suits many in Congress, who can claim alarm and oversight while doing nothing to actually constrain development. What it does not suit is anyone who believes that decisions about technology capable of threatening critical infrastructure should be made through transparent processes subject to public input and electoral accountability.
The irony is that Congress has spent the week conducting other forms of oversight that reveal how little they have learned. The House advanced "must-pass" bills with strategic sweeteners attached to appease specific voting blocs—corn-state Republicans received ethanol concessions; hardline conservatives received other unspecified favors. This is transactional governance at its most naked. Yet when it comes to the actual future of American security, Congress cannot negotiate because it lacks the leverage to do so. The tech companies hold all the cards.
The Silence Over Consequences
Two stories from the international section deserve more attention than they received. First: in the two months since the Pentagon allegedly conducted a strike on an Iranian school that killed civilians, the department has maintained silence on the investigation. "Highly unusual," according to former U.S. officials quoted by the BBC. This silence itself constitutes a kind of policy—a message to allies that consequences need not be discussed, and to adversaries that accountability no longer constrains military action. Second: the Senate rejected an effort to require congressional approval before Trump pursues military action against Cuba, underscoring that Republican control of the chamber means presidential war powers face minimal constraint.
These stories, seemingly unrelated to the AI briefings, share a common thread: the abdication of institutional oversight in favor of concentrated executive power—whether vested in the presidency or distributed among private technology firms. The pattern extends across the week's news. David Morens, a former adviser to Anthony Fauci, was indicted for allegedly circumventing FOIA requests to keep public records hidden. Meta denies that WhatsApp content moderators can read encrypted messages, while the Trump administration closes its investigation. Sony may have implemented a 30-day license verification requirement for digital game purchases that players cannot actually own in any traditional sense—and the company offers no official explanation, preferring silence to accountability.
In each case, someone with power has decided that the public does not deserve to know something. Sometimes the power is governmental, sometimes corporate. Sometimes the secrecy is justified by national security; sometimes by terms of service agreements that no one reads. The cumulative effect is the same: the relationship between institutions and the people they serve has calcified into one of mere subjects receiving whatever information their rulers believe they need.
The Briefing Unfinished
When the classified briefing on AI capabilities concluded Thursday, congressional staffers left the room presumably understanding more about the threat than they had upon entering. But understanding is not the same as agency. They understand that a thing is dangerous. They do not understand how to regulate it, whether regulation is even possible, or what the consequences will be if they get it wrong. The companies understand these things better than Congress does. And so, in the strange inversion of power that defines 2026, those who possess knowledge cede it only to those with authority, but those with authority must beg for knowledge.
This is no way to govern a democracy at the edge of transformative technological change. Yet it is, precisely, how we are governing it.