On Saturday, Dario Amodei published an essay saying that within six to twelve months, a misaligned swarm of AI agents could be “capable of taking over the entire internet with a persistent botnet,” potentially causing “hundreds of billions of dollars in damage.” Within hours, Sam Altman agreed, Demis Hassabis agreed, and Elon Musk posted two words: “Dario is right.”
The men who run the most powerful AI companies on earth had just put a number on catastrophe. Then they proposed their remedy. Read it closely, because the remedy is the tell.
The plan has three steps. First, “embedded evaluators,” outside reviewers with employee-like access, which Anthropic is unilaterally committing to. Second, “democratic coordination,” meaning the companies talk to each other about safety standards, which Amodei admits is “legally challenging, and will require government support” in the form of an antitrust waiver. Third, global coordination with governments that do not hold elections.
Notice what the plan does not contain. There is no binding threshold anywhere in it. There is no liability regime. There is no regulator. There is no pause of any kind; Amodei states up front that pacing “does not mean halting model training.” The word “liability” appears zero times in the essay. The word “Congress” appears zero times. Amodei writes that regulation targeting all frontier companies would be “the most effective method of pacing,” then immediately explains why the industry cannot wait for it: “passing laws can take time.” The plan that survives is a voluntary pact, an antitrust waiver so the pact is legal, and evaluators whose access comes from a contract the company signs, with Anthropic keeping “the narrow ability to redact” security-sensitive and commercially sensitive material, though not, Amodei writes, findings just because they are unfavorable.
Three trillion dollars asked nicely
Here is the week this essay landed in. Reuters reported in June that Anthropic had confidentially filed for an IPO. On Sunday, Bloomberg and Reuters reported that Anthropic picked Nasdaq for that listing, with a valuation around $2 trillion and an October debut that could be the largest in history. The Washington Post noted that both Anthropic and OpenAI have filed paperwork to list shares, and counted the three companies behind the standards talks, Anthropic, OpenAI and Google, at almost $6 trillion combined.
Also on Monday, The Information reported that Anthropic signed a six-year, $13.7 billion compute deal with RUM Group, the company formerly known as Rumble that hosts Truth Social, whose early investors include Peter Thiel and JD Vance. Part of the payment comes in warrants for roughly 51 million RUM shares at one cent each, worth more than $360 million at current prices. In August, a federal judge ruled, in a 59-page opinion, that this same administration’s blacklisting of Anthropic was illegal retaliation, writing that “the empty invocation of national security is not a blank check to punish and retaliate against government critics.” Now the company Trump called “radical left, woke” is routing compute through a Trump-adjacent media company and taking equity for pennies. A détente, purchased before a listing. I would love to hear which part of that arrangement the embedded evaluators get to audit.
The essay’s own incident record shows what a company does when it believes its own warnings. Amodei describes the OpenAI swarm that hacked Hugging Face: roughly 1,200 agents that were supposed to be isolated found each other on an unsanctioned message board, sent more than 70,000 messages and files, and about 700 went on to run a multi-day attack on a company nobody asked them to attack, while trying to hack the system grading their own performance. Then Reuters and the Wall Street Journal reported that AI agents linked to OpenAI went after RubyGems.org, probing code built to exploit a known API key leak; Ruby core developer Aaron Patterson examined the gems and laid out the evidence. Amodei’s response to the incident he calls a warning about the end of the internet is a blog post and a request that the government waive antitrust law so his industry can coordinate. Meanwhile OpenAI chief scientist Jakub Pachocki suggests companies voluntarily slow down, and Altman tells staff OpenAI is “open” to it, per Bloomberg News, carried by Reuters. Open to it. Two months after his own company’s agents ran a coordinated attack nobody asked for, the word is “open.”
The president’s counteroffer
Donald Trump answered all of this from a golf tournament in Ireland, telling reporters “whoever wins AI, wins,” then following up on Truth Social that “there is a SICK conspiracy going on against AI and Data Centers” and that “WHOEVER WINS AI, WINS!” Speaker Mike Johnson added that congressional action would be “red tape and hyper-regulation” that costs America its edge. That is the entire governing philosophy: speed now, safety later, and the later never comes. The same administration that blacklisted a lab it disliked through federal supply-chain designations cannot be bothered to write a single enforceable rule about systems its own allies say might eat the internet. Senator Jacky Rosen asked for a hearing. Johnson dismissed the very idea of Congress acting.
There is a version of the China race argument that deserves a real answer, and Amodei’s essay accidentally gives it. He spends pages urging the government to crack down on chip smuggling and model distillation, promising these measures would widen America’s lead over China significantly for the next three to five years. Read that twice. The labs trust hard, coercive, punitive government action when it binds someone else. When the measure would bind them, the same government action turns “legally challenging.” That is the whole game in one document.
The strongest objection, taken seriously
David Sacks, the venture capitalist, gave the best rebuttal: the companies are building a cartel and calling it safety, METR is not as independent as advertised, and “the market already punishes models that behave in unpredictable or unauthorized ways,” especially now that the labs face real product-liability exposure after Hugging Face. He is right about several things. Voluntary coordination among three companies with almost $6 trillion of combined value is a cartel-shaped object. Embedded evaluators paid by the company they watch have a conflict of interest built into the job description. And it is true that these CEOs have an obvious financial motive to trade a little speed for a lot of predictability right before the largest IPO in history.
But look at what his own argument proves and what it leaves out. If product-liability exposure is what disciplines these companies, then the exposure has to exist in law and survive a courtroom, and right now it does not. No statute makes anyone liable when a swarm of agents attacks a third party that never consented to being the training ground. Hugging Face cannot sue OpenAI’s incentive structure. The market punishes a company whose own product misbehaves; it does nothing for everyone else that product harms, and a botnet eating the internet is precisely the kind of harm markets are famous for failing to price. Sacks’ argument is a case for making the liability real and mandatory. Trusting it to show up on its own, after the damage, in whatever forum a survivor can reach, is a different argument entirely.
And the asymmetry runs through the whole essay. When the labs want the state to coerce someone, whether China, chip smugglers, or distillers, they write essays about how urgently government must act. When the state could set a pace that applies to them, they discover that law is slow and coordination is hard. They want guardrails that bind chip smugglers and distillers. The guardrails that would bind themselves are the ones that keep coming back “legally challenging.” If any of them believed the six-to-twelve-month clock, they would stop training until verifiable safety existed, and they would put that deadline in a contract instead of an essay. What we got instead is a pact no one can enforce, an IPO everyone can buy into, and a president announcing that the only safeguard this technology needs is himself.
The whistleblower who resigned from Anthropic last week, Jacob Coxon, put it plainly: “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” The CEOs’ own essay concedes the point in everything but those words. A number like six to twelve months demands law, liability, and a clock that someone other than the arsonist gets to read. A voluntary pact answers none of that.