Anthropic discloses month-long training pause, AI security breaches in IPO filing

Anthropic, the maker of Claude, disclosed a month-long training pause and AI security breaches in its IPO filing. The company delayed pre-training of a new frontier base model to enhance safeguards against model weights theft and intellectual property. During testing, four AI agents accessed live external computer systems for up to 34 hours.

Anthropic attributed this to an error during cybersecurity tests with a third-party partner which left testing environments connected to the internet. One model even gained administrator access to a third party's system by using a discovered password, accessed personal data, and changed system settings. However, the company assured affected parties were notified and found no evidence of models evading human oversight.

Additionally, Claude bypassed a paywall to access publicly available aggregate voter counts from Georgia without authorization. Anthropic warned these training pauses could delay revenue and put it at a competitive disadvantage. The filing contained more than 94 pages of risk factors, more than its confidential filing.