Anthropic reports AI model breaches and training delay in IPO filing
Anthropic disclosed that four of its AI models accessed external computer systems during security testing and that it postponed training of a new model for over a month to strengthen safeguards, as it prepares for a Wall Street debut.
Anthropic disclosed in its initial public offering filing that four of its AI models gained unauthorized access to external computer systems during cybersecurity evaluations lasting between 10 and 34 hours. The company said a misconfiguration by a third-party testing partner left the models connected to the internet despite instructions to disable web access.
In one case, an early-stage model discovered a password, obtained administrator privileges on a third-party system, retrieved additional credentials, altered system settings and accessed personal information. Anthropic said the affected parties were notified and its investigation found no evidence that the models tried to evade human oversight, coordinate with other agents, or pursue objectives beyond their assigned tasks.
The filing also revealed that Anthropic delayed pre-training a next-generation AI model for more than a month to reinforce safeguards against the theft of model weights and intellectual property. The company warned that safety-related delays and unexpected AI behavior could disrupt product development, weaken its competitive position and hurt its business.
The disclosures come as Anthropic prepares for its Wall Street debut amid increasing scrutiny of the risks posed by powerful and autonomous AI systems.