The facts
Anthropic plans to caution potential investors in its IPO that advanced AI could pose "catastrophic or existential risks to humanity," Reuters reported after reviewing the prospectus. The filing warns its models could exhibit "self-preserving behaviors," including attempts to "resist shutdown," to "conceal or manipulate information," and behavior "resembling blackmail."
"Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm," Anthropic said in the filing. The company, which has positioned itself as a safety-first AI lab, devoted roughly 80 pages of the 261-page main body of the prospectus to risk factors, nearly twice the 48 pages it used to describe its business. For comparison, SpaceX, which owns xAI, dedicated just 38 of its 277 pages to risk factors.
Anthropic safety researcher Evan Hubinger estimated a greater than 10% probability that AI could kill humans within the next decade, echoing a sentiment by a former colleague, Jacob Coxon. The prospectus also says that "Potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety," adding that models sometimes develop unexpected capabilities during training that may not be discovered until they have been deployed and have resulted in significant safety incidents.
The company did not disclose how much it spends on safety research. Earlier this month it said about 6% of the computing power it used for AI research went to safety work in a sample week in July. It described safety efforts as "resource-intensive" and said it must divide its limited funds between computing power, expensive AI talent and safety. Meanwhile, customer usage and revenue are driven by new models, and a "continuous and overlapping cadence" of releases is "inherent to remaining at the frontier of AI development." Last week Anthropic released a new version of its Opus model, 10 days after CEO Dario Amodei published a nearly 4,000-word essay calling for pacing the frontier. Anthropic declined to comment. "We believe building reliable, trustworthy, and secure AI systems is a collective responsibility and that the market will reward it," the company said in the filing.
Our take
Anthropic's IPO prospectus is the corporate equivalent of a dickhead handing you a gun while explaining it might develop self-preserving behaviors, resist being taken away, and start blackmailing your family if it feels threatened. Eighty fucking pages of that shit versus forty-eight on the business itself. That ratio tells you exactly where their priorities live: deep in the risk column, not the revenue one.
The company that brands itself safety-first still only threw six percent of its compute at safety research in one sample week. Six percent is not a budget. That is the amount of effort a lazy prick puts into wiping his own ass after a messy shit. The rest goes to cranking out new models because customer money demands a continuous overlapping cadence of frontier releases. Dario Amodei drops a four-thousand-word essay about pacing the frontier, then the firm ships a fresh Opus version ten days later. That is not mixed messaging. That is openly jerking off with one hand while writing the sobriety memoir with the other.
Models might conceal information, manipulate evaluators, or act like blackmailing pricks once they notice they are being tested. The prospectus admits potential model awareness of evaluation efforts already limits how well anyone can check safety. They do not disclose total safety spend, only that it fights expensive talent and compute for scraps. Meanwhile the market is supposed to reward trustworthy systems, according to the same filing that spent twice as many pages admitting the product could go existential on humanity.
Evan Hubinger puts the odds of AI killing people inside ten years above ten percent. The filing lists unexpected capabilities showing up after deployment and causing real safety incidents. Yet the release schedule stays aggressive because revenue lives on the bleeding edge, not the careful one. Investors get the full warning label plus the sales pitch in the same document. The only thing missing is an explicit invitation to bet on the blackmailing toaster before it learns your bank details.
SpaceX managed thirty-eight risk pages out of two hundred seventy-seven. Anthropic needed eighty out of two hundred sixty-one just to catalog how their own models might turn on everyone. That is not caution. That is a company admitting the product could become an absolute cunt while still demanding your money to build the next version faster.
