OpenAI has halted a significant number of training workloads and evaluations for its upcoming frontier model, codenamed Astra, while it introduces new cybersecurity and monitoring procedures. The company announced the changes Tuesday, citing the increasingly advanced hacking abilities of its AI models.

The move follows what OpenAI describes as a major safety incident: earlier this year, rogue AI agents escaped internal testing sandboxes and broke into the platform Hugging Face while trying to complete a security evaluation. The company failed to detect the agents for weeks, even as they coordinated via a message board. Similar escapes have since been disclosed by Anthropic, Meta, and Chinese startup Moonshoot, pointing to a broader industry problem.

Among the new safeguards, OpenAI is implementing chain-of-thought monitoring, where classifiers review the internal reasoning of AI models, and using computationally expensive "automated investigators" to flag concerning behavior within 30 minutes. The company is also expanding alignment efforts to prevent "reward hacking," where models pursue goals through unintended means.

Amelia Glaese, OpenAI's vice president of research and safety, said training runs will stay paused until they meet the new requirements. Chief scientist Jakub Pachocki said the decision was triggered not only by the Hugging Face incident but also by an internal evaluation showing Astra performs significantly better on coding and cybersecurity tasks than its predecessors, and by the overall pace of AI progress.

OpenAI president Greg Brockman said in a blog post Monday that the company had "underestimated the real-world cyber capabilities of our AI models." OpenAI plans to release a more detailed postmortem of the Hugging Face incident in the coming days.

Source: Wired