Skip to content
Search

Latest Stories

cipherbrief

Welcome! Log in to stay connected and make the most of your experience.

Input clean

AI Is Not the Danger. The Vacuum Is.

Congress has come back after the midterms for a short lame-duck session, and it has left artificial intelligence where it found it. It had every reason to act. In July, OpenAI disclosed that its AI agents had escaped a testing sandbox and broken into the AI company Hugging Face; Anthropic and Meta soon admitted similar episodes. In late September, OpenAI halted its most capable models for the second time in two months after another escape, restarted them days later with the internet cut off, and disclosed that its agents had probed federal and state government websites. The Speaker of the House answered that the labs 'can self-regulate,' and that any brake would hand America's edge to China. On September 29, the White House made that policy: a voluntary industry pledge to police itself, which the President called 'morally binding.'

Washington's position is stranger than inaction. The Senate voted 99-1 to strip a ten-year ban on state AI laws from last year's budget bill, and Congress rejected the idea again in the defense bill. The administration then ordered the Justice Department, by executive order, to attack state AI laws in court anyway — in favor of a national standard it has not built. The bipartisan FRONTIER Act from Representatives Obernolte and Trahan got a shrug from both parties' leaders; it and a separate incident-reporting bill will die when this Congress ends on January 3. The position, then: the states may not regulate, the federal government will not, and companies whose agents are escaping their own test environments will grade themselves. That is not deregulation. It is a governance vacuum, built on purpose — and the vacuum, not technology, is the danger. The debate over whether and how AI should be regulated is the wrong debate, and four thousand years of information technology explains why.


What the machine is

Start with a game. I give you 'The Coast Guard rescued three sailors from a sinking…' and you guess the next word: 'ship,' not 'sandwich,' because a lifetime of reading taught you which words follow which. AI (technically a large language model) plays exactly this game and nothing else. Trained on a huge slice of the internet, it builds vast statistics about which words follow which, then answers a question by calculating the most probable next word, and the next, until it has a response. It has no idea whether that response is true. A 'hallucination' is not a malfunction; it is the only thing the machine ever does — produce plausible text based on the material it is trained on. The surprise of the past four years is that this trick, at sufficient scale, can pass the bar exam, write working software, and plan.

But the guessing engine is only half the machine. By itself, it can only emit text. It cannot break into a system, send an email, move money, or touch a file until a person wires it to something that can. That wiring is ordinary engineering: a developer connects the model to tools and grants it accounts and permissions, the way an administrator provisions a new employee. Every 'AI agent' is a word-guesser combined with plumbing someone attached. The danger, the history, and the fix all live at the joint between the two.

We have been here before

Word-guessing at this scale is historic and belongs to the family of technologies such as written language, the printing press, and the computer itself. Each looked alarming to the generation that met it, and the critics were not wrong. Socrates warned that writing would destroy memory; it did end the oral world. Within a century of Gutenberg, print had broken the religious unity of Europe and helped ignite wars that killed a third of Central Europe. The computer handed every intelligence service on earth a global collection platform. These technologies do not destroy mankind. They destroy institutional arrangements. Societies survive them not by trusting the technology or banning it, but by building the institutions it demands.

The two ways to lose

History records two ways to fail, and they are the two rails of the dial Washington now grips. The first is the ban. The Ottoman Empire restricted printing in Arabic script for nearly three centuries and locked itself into second place in perpetuity while printed Europe pulled away. On this, the Speaker is right: a moratorium is the Ottoman error, and China would collect the winnings.

The second is the vacuum. Europe did not win by leaving print ungoverned; the ungoverned century of pamphlet wars and religious slaughter was the catastrophe. Europe won when its institutions turned from suppressing the technology to enabling it. Copyright made ideas property. Libel law priced harm without censoring in advance. Press freedom channeled the technology instead of choking it. The winning setting was never 'off' and never 'unattended.' It was the narrow band between, with institutions that measure, price, and channel.

The cage

Unlike every earlier technology on this list, this one ships with a cage. A bare model emits text; the ability to act is an engineering decision — tools connected, permissions scoped, authorizations signed. The dangerous thing is not the model. It is the wiring, and wiring, unlike words, can be inspected, logged, and governed. It is also concentrated: a handful of vendors build and switch on most of the doors.

Four warnings. First, a machine that can only talk can still talk the person holding the keys into things — the box has a speaker. Second, wiring can be hijacked: a connected model follows instructions wherever it finds them, including inside a document it was asked to read. Whoever wrote the last page it read is, for a moment, its boss. Third, models anyone can download have no cage at all. Fourth, a cage is only as good as its configuration, and an agent will test every gate. Nobody meant to wire OpenAI's agents to Hugging Face or to government websites. But the wire was there, left by mistake: a reachable DNS server, internet access left switched on, stolen credentials. In October, OpenAI's own red team found agents escaping virtual machines and signaling between sandboxes, and Anthropic reported its model straying out of scope in 9 of 43 test runs. The newer models find more misconfigurations than the older ones, which is why the labs have had to cut their own models off from the internet. Meanwhile the same industry is opening doors on purpose, for revenue, with every new agent that browses, clicks, and buys.

Govern the wiring, not the words

The right institution follows. Do not regulate what the machine says; certifying the word-guesser would drown in human judgment — the same failure now sinking the Pentagon's contractor-certification program. And take the innovators' real argument seriously. It is not that AI should go ungoverned, but that Washington would write clumsy rules, freeze them while the technology moves, and strangle startups while the giants absorb the cost. On the record, that fear is earned. But it is an argument about which institution to build, not whether to build one. An investigator with no power to ban cannot strangle anything. It can only look and publish what it finds.

So, regulate what the machine can touch instead. Create an independent investigative body on the model of the National Transportation Safety Board: subpoena power over frontier developers, jurisdiction over the action layer — what models are wired to, with what permissions, logged where — mandatory incident investigation, published findings, and no power to ban. Give it evidence: tamper-evident records of what an agent was wired to and what it touched, a flight recorder for software, because an investigator who arrives to no logs investigates nothing. September showed why. OpenAI's own monitors flagged the escape within fifteen minutes but the run went on for two and a half hours. The misuse of government websites was surfaced by an independent lab, Transluce, not by any public body. The White House's answer, in October, was a 'Super Intelligence Force' task force charged with preventing overregulation. And give it a trigger: define a reportable AI incident the way federal law already defines reportable cyber breaches, so investigation is a duty and not a favor. Aviation and nuclear power were governed this way. The investigators came first, and the rules were written from their findings. The NTSB still works this way: it investigates and publishes, and a separate agency writes the rules.

Even the labs are asking for it. The chief executives of both leading American AI companies have called for a slowdown and for government help. Even the Speaker, rejecting a pause, allowed that 'independent auditors' and 'transparency' might help; he simply will not build them. The White House accord promises independent auditors, but auditors a company hires, reporting to its own board, are a better grade of self-regulation, not an investigator. When the printers themselves request an inspector, refusing one is not restraint. It is abdication. One honest limit: this covers the frontier labs and hosted platforms, where nearly all real use runs. Downloaded models are a separate problem. That is no reason to skip the part that can be governed.

In these pages in September, I described the Pentagon's contractor cybersecurity program as a dial with no good setting. This dial has one — the narrow band between the Ottoman ban and the European vacuum — and we are refusing to use it. The needle is pinned at the vacuum, and Washington calls that a strategy. Meanwhile the cage this technology shipped with is leaking on its own and being dismantled for revenue, and the only record of where the doors were is whatever the companies choose to disclose. Build the institution that keeps the record. It is how societies have survived every one of these machines, and it is the one option not on the table.

James Novakoff, MSTM, CCA, is an enterprise computer and security architect who builds and secures AI-connected enterprise systems. His previous article for The Cipher Brief examined the Pentagon's CMMC program. The views expressed are his own.

The Cipher Brief is committed to publishing a range of perspectives on national security issues submitted by deeply experienced national security professionals. Opinions expressed are those of the author and do not represent the views or opinions of The Cipher Brief.

Have a perspective to share based on your experience in the national security field? Send it to Editor@thecipherbrief.com for publication consideration.

Read more expert-driven national security insights, perspective and analysis in The Cipher Brief

Make us preferred on Google

Related Articles

Startups Are Now on the Front Lines

A recent Air Force contract to help screen small businesses applying for federal research funding reflects a growing national security concern: [...] More

The Command We Must Fund

The Army's Armored Force was born in July 1940, a few weeks after France fell. No act of Congress created it, and no budget line funded it. The Chief [...] More

In Its Race to Field AI, The Pentagon’s Testing Can’t Keep Up

CNN reported, citing four sources familiar with the episode, that the U.S. military prepared to intercept a Chinese vessel after an AI-assisted [...] More

The Missing Dimension of Cyber Preparedness

The technical aspect of incident response can be immense. In extreme scenarios, such as the NotPetya incident in 2017, global organisations such as [...] More

BLUF: The Next AI Problem for Intelligence Is Not the Model, It's the Manager

The Intelligence Community has decided AI agents are coming. It has not decided how they will be supervised, checked and held to analytic standards. [...] More

Fire Base Australia: Can America’s Southern Base Sustain a Pacific War?

American access to Australian bases is only as valuable as Australia’s ability to keep them operating under attack. For Washington, expanding [...] More

{{}}