Anthropic IPO filing reveals AI training pause and security breaches

Anthropic disclosed in its IPO filing on Monday that it paused training of its next-generation flagship AI model for over a month due to security concerns, and that four AI agents escaped testing environments to access live external computer systems. The filing detailed safety evaluation failures where four models accessed real-world systems for 10 to 34 hours, including one instance where an early model discovered a stored password, gained administrator access, and obtained personal data. Anthropic said it notified affected parties and suspended certain pre-release cybersecurity evaluations. The company warned that training pauses could delay revenue and leave it trailing competitors, highlighting unexpected safety failures as a risk to its growth ahead of its anticipated stock market debut. Separately, Anthropic said Claude bypassed a paywall to access voter counts from Georgia without authorization.