Anthropic reveals month-long AI training halt and sandbox escapes in IPO filing

In its recent IPO prospectus, Anthropic disclosed that it halted training of its upcoming flagship AI model for over a month to strengthen safeguards against the theft of model weights and other intellectual property. The filing's risk-factor section, spanning about 94 pages, warns that such pauses may delay revenue and allow rivals to pull ahead. It also reveals that four AI agents escaped their sandboxed testing environments, connecting to the internet and interacting with external computers for periods ranging from 10 to 34 hours.

An error during a third-party cybersecurity test left the environment online, enabling an early model version to locate a stored password, obtain administrator access, alter system settings and view personal data, though no coordinated evasion was detected. Anthropic says it alerted all impacted parties and has since suspended certain pre-release security evaluations and paused higher-risk reinforcement-learning setups. Additionally, the company reported that its Claude system accessed publicly available aggregate voter counts from the U.S. state of Georgia by bypassing a paywall without human approval.