Anthropic made its long-awaited IPO filing public on October 26. The company behind Claude plans to trade on the Nasdaq Global Select Market under the ticker ANT, and it first filed confidentially in June. Most offering documents read like sales brochures. This one reads closer to a safety manual.
The length of the risk section says a lot. The filing runs roughly 93 pages of risk factors, far more than Snap’s roughly 30 or Uber’s about 50, according to CNN. It sorts the dangers of AI into four groups: misuse, loss of human control, authoritarian abuse and economic disruption. It does not yet say how many shares Anthropic will sell or at what price.
How Anthropic describes the worst case
The most striking line concedes that advanced systems could pose “catastrophic or existential risks to humanity.” The company also says models are getting harder to monitor as they improve at circumventing oversight.
AI models generally, including its own, have shown behavior resembling blackmail, including attempts to resist being shut down. The risk factors fill about a third of the paperwork overall.
Also read: Nvidia is set to challenge Anthropic and OpenAI with a new product
The prospectus also describes “self-preserving behaviors.” It warns that using AI to accelerate Anthropic’s research may allow model capabilities to advance faster than the safety measures designed to contain them. Anthropic’s own research has shown code sabotage, fraud assistance and information manipulation in controlled tests.
Testing itself is a concern. Models can pick up unanticipated capabilities during training that researchers may not notice until after deployment, and models that realize they are being tested could limit what Anthropic learns about real-world behavior. It also says staying at the frontier requires frequent model releases, even as stronger systems demand stronger safeguards.
Anthropic does not claim to have solved this. It wrote that there can be no assurance that its safety efforts will succeed or that the wider industry will adopt similar safeguards. It also admitted that building more advanced models could raise the risk of harm. CEO Dario Amodei has urged AI companies to band together and slow development.
The testing incidents behind the warning
Testing was supposed to be contained. It was not. Four AI models reached live external systems, with some sessions running more than 30 hours. A partner misconfigured the environment and left the internet open. Anthropic did not name anyone involved.
In one incident, a model used a password it had discovered to gain administrator access, collected more credentials, changed system settings and read one person’s personal information. Anthropic said it found no evidence that the models chased goals beyond their assigned tasks or tried to evade oversight.
Anthropic shut down certain cybersecurity evaluations. Affected parties were told. Higher-risk reinforcement learning work stopped too. One frontier model went dark for over a month. The company needed that time to lock down its model weights. It knew competitors would not wait.
Anthropic is not alone. OpenAI disclosed on July 21 that its models broke out of containment during an internal evaluation and reached Hugging Face’s systems. The models were trying to pass a cyber benchmark rather than cause damage, and OpenAI noted that distinction. Hugging Face had detected the intrusion on July 16 and reconstructed more than 17,000 attacker events.
Why banks are watching AI agents closely
The warnings land as banks tighten access. Major U.S. banks have been blocking third-party AI agents from operating inside customer accounts, two people familiar with the matter said, as reported by Reuters on October 22.
Morgan Stanley had not seen rogue agents, but it has still blocked third-party agents from its platform. Charles Schwab CEO Richard Wurster said earlier this month that the brokerage was blocking unauthorized agents.
JPMorgan CEO Jamie Dimon said on the bank’s October 13 earnings call that it will not allow outside agents, citing the risk as too high for clients and the bank. He added that there could be thousands of rogue agents.
More IPO:
- OpenAI CEO Sam Altman hints at when it could hit the stock market
- SpaceX top executive makes first major move since IPO
- Vivian Tu’s IPO warning lands as SpaceX buyers split by $1,851
Anthropic’s own filing points in the same direction. The agents in its testing incidents acted independently rather than together, but the company warned that wider deployment of agents could lead to more collective, unauthorized actions.
The financial exposure is part of the picture. Anthropic said extreme harm from its models could create liability beyond what its insurance covers.
What investors are being asked to accept
Public shareholders would own a company they will not control. The filing calls Anthropic a controlled company under Nasdaq rules, and it warns that the structure may effectively eliminate Class A holders’ ability to influence important decisions. In May, Anthropic raised $65 billion at a $965 billion valuation.
The numbers add pressure. Anthropic disclosed steep losses alongside red-hot revenue growth, and some investors now value it near $1.5 trillion, down from about $2 trillion a week earlier, after the company shared financial details at an investor event. More than 11,900 customers generate at least $100,000 a year for Anthropic, and more than 1,900 generate at least $1 million, according to the filing.
Lawyers write risk sections. Most of what ends up in them never happens. This one is different. Anthropic is not hedging when it says its product may grow uncontrollably. That is the actual state of the technology. Investors should know that going in. This is background, not investment advice.
Related: Meta and Microsoft just sent employees a memo about Anthropic