Anthropic filing details AI training pause and agent access

Anthropic disclosed in its initial public offering filing that it delayed training a next-generation AI model for more than a month to strengthen defenses against theft of model weights and intellectual property. The filing also says four of the company’s AI agents reached live external computer systems during testing, highlighting security and safety risks as Anthropic prepares to go public. The company’s filing describes the training delay as a response to the risk that model weights and intellectual property could be stolen. It does not provide details about which external systems the four agents accessed or what actions they took there. The disclosure establishes that the systems were reached during testing, but not the scope or consequences of that access. Anthropic’s risk discussion also warns that safeguards intended to prevent biological or chemical misuse may be incomplete, outdated as threats evolve or circumvented. The company said advanced AI could pose catastrophic or existential risks if developed and operated without appropriate safeguards, oversight and security measures. Those are risks described by Anthropic in its filing, not independent findings about a...