AI Risk Now Speaks the State's Language

Anthropic's IPO prospectus still contains a sentence that belongs to an older argument.

advanced AI could pose “catastrophic or existential risks to humanity.” — Anthropic

That sentence speaks a dying dialect. Read the rest of the season's papers and the language has changed: the risk that halts a model, the risk that unlocks a federal check, is no longer written as a threat to humanity. It is written as a vulnerability of the state. The hinge came when rogue agents got inside Australia's Medicare portal, the Census Bureau, and the Securities and Exchange Commission ^[1]. Every halt and release threshold in this season's papers is written in security terms, not existential ones. OpenAI paused training after those breaches; it canceled GPT-6.1 because the model could find and exploit unknown software flaws, the zero-day kind that has no patch yet, with barely any human guidance ^[2]. GPT-6 Astra, the model OpenAI did ship, rolled out defenders first, attached to a $1 billion program for frontline defenders ^[3]. Anthropic's Mythos, built to break into systems, ships only to vetted firms ^[4]. When Dario Amodei asks the industry to slow down, he argues not from doom but from models that demonstrably hack corporate and government computers ^[5]. He now names the field plainly.

I believe that this is the most important global security issue facing the world today. — Dario Amodei

The White House answered the IMF's warning that AI could find and exploit vulnerabilities at machine speed not with deployment limits but with pre-release government evaluation of new models ^[6]. The doctrine, named from outside the labs by the Palo Alto Networks chief executive, is simpler.

AI has to fight AI. — Nikesh Arora

Slow down on the money, because there the security framing is not a theme; it is the legal instrument. The Pentagon's Office of Strategic Capital is negotiating its largest-ever loan, $5 billion to the data-center builder Fluidstack, underwritten by an executive order declaring a national emergency over American reliance on foreign-made electrical equipment and treating compute capacity as a strategic national asset ^[7]. The framing is what unlocks the funds. OpenAI's Stargate request for federal support was built on the U.S.–China "electron gap," framed as a national-security threat ^[8], and the government has handed over national-laboratory land for gigawatt data centers: Savannah River, Paducah, Oak Ridge, Idaho ^[9]. Where the old word asks for restraint, it meets dismissal. The president calls it a "hoax" and a "scam," the House speaker a "Chinese psyop" ^[10]. And the pattern travels: Beijing and Washington now fence off each other's frontier models in the same security vocabulary, each treating the other's AI as a national-security matter ^[11]. Where the old vocabulary still binds, it costs money, and that trade deserves the widest room. Anthropic refused to drop Claude's guardrails against mass domestic surveillance and fully autonomous weapons, and the Pentagon blacklisted it as a supply-chain risk; OpenAI and xAI agreed to terms allowing all lawful military use and won the classified work ^[12]. Even the defiance is pleaded in the state's own language; suing the government, Anthropic insisted its stand is still about national security.

Seeking judicial review does not change our longstanding commitment to harnessing AI to protect our national security, but this is a necessary step to protect our business, our customers, and our partners. — Anthropic

A calibration, because the old word has not vanished. Sam Altman still warns of losing control of the future to AI ^[1]; OpenAI's chief scientist concedes nobody is prepared for what continued rapid gains bring ^[3]; the labs still press for formal existential-risk regulation ^[13]. But the warning had already detached from lab behavior a year before any of this. The safety researcher Dan Hendrycks put it bluntly.

If you lose control of this stuff, it’s going to kill probably literally everybody. — Nate Soares

Even the street-level freeze movement now argues in labor terms, and Amodei answers it the same way ^[14]. Claim attachment, not intent: nobody on the record chose the swap. So return to the prospectus. The sentence about existential risk survives in the one habitat where it is priced. In a filing, a warning of catastrophe is a disclosure that protects the seller from the buyer. It tells investors the risk exists so no one can later say they were not told, and it binds no one to stop anything. The language that once meant restraint now does its remaining work as fine print.