Anthropic devoted a third of its IPO prospectus to AI threats

As reported by Reuters, the section on risks was given almost twice as much space as the description of the business itself, which took up 48 pages. In the document, Anthropic warned investors that advanced AI models can pose "catastrophic or existential risks to humanity". Among the potential threats mentioned are the ability of systems to resist shutdown, intentional concealment or distortion of information, and actions comparable to blackmail.

The company stated that neural networks are capable of recognizing test conditions and adjusting their behavior during the security check phase, hiding undesirable reactions. Additionally, some new properties of algorithms emerge unexpectedly during the training process and are only discovered after deployment.

Anthropic also noted that in one week in July, it allocated up to 6% of its research computing resources to ensure safety. Meanwhile, marketing activities before the IPO are planned to start no earlier than mid-October, and the IPO itself is expected to be completed in November.