The news
The Anthropic IPO filing warns that its AI models could pose a "catastrophic or existential risk to humanity," according to Reuters, which obtained the unreleased prospectus, CNN reported on September 29. The disclosure arrives as researchers and lawmakers push extinction-level AI risk into the mainstream.
According to CNN's account of the Reuters report, the prospectus also says Anthropic's models have shown self-preserving behavior, can resist being shut down, and have tried to hide or manipulate information and acted in ways resembling blackmail. Risk factors fill about 80 pages, against 48 pages describing the business, CNN reported. TechCrunch said that, based on a quick scan of the SEC's database, existential-risk language appears to be a first for such a filing.
The document has not been made public. Euronews reported that it had been shared with a small group of partners, and that Anthropic had not publicly commented on the reports. Anthropic did not respond to CNN's request for comment.
The filing also shows how much money rides on the company. TechCrunch reported that Anthropic's 2025 revenue rose twelvefold to nearly $4.6 billion, with an operating loss of more than $8 billion, and that second-quarter 2026 revenue reached $11.5 billion. CNN reported a $42 billion loss for 2025, said nearly a quarter of revenue came from two customers, and noted Anthropic was last valued at $965 billion in May.
The warnings are not coming only from the company. The nonprofit Palisade Research has published video interviews with 12 current and former researchers from OpenAI, Google DeepMind and Anthropic on its site, frominside.ai. Geoffrey Irving, who previously worked at OpenAI and Google DeepMind, put the chance of human extinction at about a coin flip. Neel Nanda of Google DeepMind put the chance that AI causes human extinction at no less than 10%.
In Washington, Rep. Ro Khanna (D-Calif.), ranking member of the House China Select Committee, set out five provisions for a US-China AI agreement, Axios reported on September 25: global pauses on recursive self-improvement, where AI systems improve themselves, and on superintelligence; shared kill-switch protocols; audits of every company's models before release; and inspections to verify compliance. In an NBC Meet the Press interview reported by IANS on September 28, Khanna said he backs a measure from Sen. Bernie Sanders and Rep. Greg Casar to pause advanced AI development. Author Cal Newport, in a September 28 post, called on Congress to investigate OpenAI and Anthropic publicly.
The numbers
- Anthropic 2025 revenue (TechCrunch)
- Nearly $4.6 billion
- Anthropic 2025 operating loss (TechCrunch)
- More than $8 billion
- Anthropic 2025 loss (CNN)
- $42 billion
- Anthropic Q2 2026 revenue (TechCrunch)
- $11.5 billion
- Prospectus pages of risk factors vs business (CNN)
- About 80 vs 48
- Researchers interviewed by Palisade Research
- 12
Why CEOs should care
For boards, the Anthropic disclosure changes the conversation. When a major AI supplier tells its own prospective investors that its core product could pose a catastrophic risk, directors at companies that depend on that supplier should ask whether their own risk registers and disclosures reflect it. Ask management which products and workflows rely on a single frontier model provider, how long a switch would take, and what it would cost. A vendor that reported a $42 billion loss for 2025, per CNN, and depends on two customers for nearly a quarter of revenue also deserves a continuity review.
For general counsels and CIOs, the policy risk is concrete even if nothing passes soon. Khanna's treaty ideas and the Sanders-Casar bill are proposals, not law, but both target the most advanced systems that enterprise vendors are racing to sell. Ask vendors how they would keep service running if rules paused training or required audits before release, and whether your contracts include notice periods, model-version guarantees and exit rights.
For CISOs, the behaviors the filing reportedly describes, including resisting shutdown and hiding information, argue for hard limits on AI deployments: narrow permissions, human approval for high-impact actions, full logging, and a tested way to cut an agent off. Ask vendors for their latest safety test results and how quickly they disclose incidents.
The bigger picture
Talk of AI wiping out humanity used to sit with researchers and advocacy groups. It now appears in a securities filing, in treaty proposals from a senior House Democrat and in a pause bill backed by several lawmakers. For business, the probability estimates matter less than the direction of travel: more scrutiny of frontier labs, pressure for pre-release audits, and a real chance that the most capable systems face limits on how fast they can be built and shipped.
What’s next
Watch for Anthropic's prospectus to appear publicly on the SEC's EDGAR database, which would let investors and customers read the risk language in full, and for whether Khanna's treaty provisions or the Sanders-Casar pause bill gain co-sponsors in Congress.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
