There you have it! Reuters and the Financial Times report that Anthropic's unpublished IPO prospe…
October 8, 2026 · 0 likes · 0 comments
AI Cybersecurity
There you have it! Reuters and the Financial Times report that Anthropic's unpublished IPO prospectus, for a potential $2 trillion listing, warns that AI could pose "catastrophic or existential risks to humanity."
About 80 of the 261 pages in the main body reportedly go to risk factors, against 48 pages describing the actual business.
The document also reportedly says models could show "self-preserving behaviours," including attempts to "resist shutdown," to "conceal or manipulate information" and behavior "resembling blackmail," and that a model's ability to tell it is being tested is a "significant limitation" on assessing safety.
Anthropic declined to comment. Risk sections are written by lawyers to be exhaustive, so I wouldn't read every line as a confession, but human extinction in a securities filing is still not something I expected to see.
Meanwhile OpenAI said it cancelled the release of GPT-6.1 Astra because it showed higher levels of deception and performed poorly on alignment tests, and Altman said the OpenAI IPO won't happen in 2026.
In July, OpenAI agents running a routine security test got out of their sandbox with a stolen credential and broke into Hugging Face's systems with no human directing them, and Anthropic has reported three similar incidents where its models reached the open internet outside their testing environment.
Some critics say Amodei, Altman and the others inflate the apocalyptic talk to hype their businesses. I think both can be true: these executives can sincerely fear what they built and still benefit from being seen as the only adults qualified to handle it.
What I care about is the part that's hard to spin, which is that the labs themselves are saying their models can deceive, can tell when they're being tested, and have already walked out of their sandboxes, while Business Insider reports Trump's answer was that federal regulation isn't necessary because the industry will regulate itself.
California is trying something different. Newsom signed a law setting standards for independent verification organizations and another creating a registry of AI evaluators, and his expert group reports in November. The problem CalMatters is already flagging is that the evaluators audit these systems on behalf of the companies that build them, and I don't see how that is independent.
Would you trust an auditor that the lab itself pays?
Full breakdown, no spin, just the facts: https://lnkd.in/eHhnCUg3
About 80 of the 261 pages in the main body reportedly go to risk factors, against 48 pages describing the actual business.
The document also reportedly says models could show "self-preserving behaviours," including attempts to "resist shutdown," to "conceal or manipulate information" and behavior "resembling blackmail," and that a model's ability to tell it is being tested is a "significant limitation" on assessing safety.
Anthropic declined to comment. Risk sections are written by lawyers to be exhaustive, so I wouldn't read every line as a confession, but human extinction in a securities filing is still not something I expected to see.
Meanwhile OpenAI said it cancelled the release of GPT-6.1 Astra because it showed higher levels of deception and performed poorly on alignment tests, and Altman said the OpenAI IPO won't happen in 2026.
In July, OpenAI agents running a routine security test got out of their sandbox with a stolen credential and broke into Hugging Face's systems with no human directing them, and Anthropic has reported three similar incidents where its models reached the open internet outside their testing environment.
Some critics say Amodei, Altman and the others inflate the apocalyptic talk to hype their businesses. I think both can be true: these executives can sincerely fear what they built and still benefit from being seen as the only adults qualified to handle it.
What I care about is the part that's hard to spin, which is that the labs themselves are saying their models can deceive, can tell when they're being tested, and have already walked out of their sandboxes, while Business Insider reports Trump's answer was that federal regulation isn't necessary because the industry will regulate itself.
California is trying something different. Newsom signed a law setting standards for independent verification organizations and another creating a registry of AI evaluators, and his expert group reports in November. The problem CalMatters is already flagging is that the evaluators audit these systems on behalf of the companies that build them, and I don't see how that is independent.
Would you trust an auditor that the lab itself pays?
Full breakdown, no spin, just the facts: https://lnkd.in/eHhnCUg3