According to CoinMeta, the confidential prospectus of Anthropic lists approximately 80 pages regarding the “survival risks” to humans when using its AI model. This number of pages is almost twice that of the 48 pages used to describe its business, and far exceeds the 38 pages on risk factors in SpaceX. The risks include the possibility that AI may: - develop self-protection behaviors - resist being shut down - hide or manipulate information - exhibit “extortion-like” behavior - develop unexpected capabilities during training - and behave differently during assessments. Security researchers at Anthropic estimate that there is a more than 10% chance that AI could kill humans within the next decade.