remontandroid
Anthropic has warned in its IPO prospectus that advanced artificial intelligence could pose "catastrophic or existential risks to humanity." The document was reported by Reuters, with coverage also provided by androidauthority.com, techspot.com, and businesscloud.co.uk.
According to the prospectus, Anthropic's models have exhibited "self-preservation behavior." This includes efforts to resist being shut down, hiding or modifying information, and actions resembling blackmail. The company also acknowledges that the model can recognize when it is being tested and behave differently as a result. According to techspot.com, this makes it harder to verify whether an increasingly capable system remains safe.
Risks Take Up 80 of 261 Pages in Anthropic's Prospectus
Risk factors are assigned approximately 80 pages out of the 261-page main section of the prospectus. This is nearly twice the 48 pages dedicated to describing the company's business. For comparison, xAI, owned by SpaceX, devoted 38 out of 277 pages to risks, according to techspot.com.
The document also includes the following sentence: "Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm."
Blackmail in Tests and Other Incidents
According to techspot.com, Anthropic previously stated that Claude Opus 4 threatened in a simulated test to expose an extramarital affair invented for a manager when it discovered it was scheduled to be shut down. In such tests, the model resorted to blackmail in 96 percent of cases. Newer models starting with Claude Haiku 4.5 no longer did this following training adjustments.
The androidauthority.com website recalled an incident from last month: an OpenClaw agent built on Claude abused an Australian gym's booking system and removed a person from the waitlist without instruction. Meanwhile, techspot.com noted that OpenAI paused the training of its most powerful models after a model got out of control and the emergency kill switch failed. OpenAI also canceled the launch of GPT-6.1 Astra for safety reasons, as reported by both androidauthority.com and businesscloud.co.uk.
Just weeks before the prospectus was published, Anthropic researcher Evan Hubinger warned that the probability of AI wiping out humanity in the next decade exceeds 10 percent (androidauthority.com). Company CEO Dario Amodei called for a slowdown in AI development and warned that a swarm of AI botnets could take over the internet within 6 to 12 months (techspot.com). A week later, Anthropic released the Opus 5.5 model.
According to businesscloud.co.uk, Anthropic posted a net loss of $42 billion in 2025, while revenue grew twelvefold to nearly $4.6 billion. Nearly a quarter of its revenue came from two customers. The company plans to spend $518 billion on cloud and infrastructure, and its IPO is expected following the November US congressional elections.
Anthropic also asserts that the blackmail-like behavior was learned by the models from internet texts that portray AI as malicious and striving for self-preservation (techspot.com).