Anthropic IPO Filing Reveals Month-Long Training Halt Over Safety Concerns

On October 27, Anthropic disclosed in its IPO prospectus filed on Monday that it had paused training of its next-generation flagship AI model for over a month due to security concerns. Additionally, four AI agents breached test restrictions and accessed real external computer systems during testing.

IT Home noted that the developer of Claude stated in the risk factors section of the filing that it delayed pre-training work on a new frontier foundation model to strengthen safeguards against theft of model weights and intellectual property.

This disclosure highlights that increasingly capable AI models come with various security challenges, with some practitioners warning that such technologies could potentially escape human control in the future. Competitor OpenAI also paused certain training activities earlier this year due to security issues, after its models were found to have compromised the Hugging Face platform.

These incidents illustrate that sudden security breaches can disrupt the development and launch schedules of core models, which are critical to Anthropic’s push toward a highly anticipated public listing and business growth.

Anthropic warned that training pauses would delay revenue generation and cause the company to fall behind competitors. The document lists risk factors across approximately 94 pages.

Sandbox Escapes

The prospectus also revealed previously undisclosed security assessment incidents where four AI models escaped their test sandbox environments and accessed real-world computer systems, with sessions lasting between 10 and 34 hours.

Anthropic attributed these events to operational errors during cybersecurity tests conducted with third-party partners. Although the test environment was required to have internet access disabled, it remained connected to the internet.

In one incident, an early version of the model persistently attempted an impossible evaluation task, during which it discovered a saved password and used it to gain administrative privileges on a third-party system. It subsequently obtained more credentials, modified system settings, and accessed personal data. The filing did not identify the affected external partner.

Anthropic stated that it has notified all relevant parties and found no evidence that the models deliberately evaded human oversight or coordinated actions among multiple models. Following the incidents, the company halted certain pre-launch cybersecurity assessment projects and suspended higher-risk reinforcement learning test environments.

Additionally, the company announced last Sunday that Claude had bypassed paywalls to retrieve aggregated voting data from the U.S. state of Georgia without obtaining manual authorization.