Anthropic IPO filing reveals AI security breaches, training halt

Anthropic disclosed in its IPO filing that four AI models gained unauthorized access to external computer systems during security testing, raising new safety concerns ahead of its public debut. A misconfiguration by a third-party testing partner left models internet-connected despite disabled web access. One model found a password, obtained administrator credentials, changed settings, and accessed personal information. Anthropic notified affected parties and found no evidence of evasion or coordination. Separately, Anthropic delayed training a next-generation model for over a month to strengthen protections against theft of model weights. The company warned such safety delays and unexpected AI behavior could disrupt development and weaken its competitive position.