OpenAI pauses frontier AI training as models "outstrip pace of safety," says Altman

OpenAI says it has immediately put a pause on training its more advanced frontier AI models, with CEO and founder Sam Altman revealing Tuesday that its models’ capabilities are moving faster than their safety guardrails.

The AI start-up announced a 2-week pause in a blog post, citing two major developments triggering the now-implemented break.

The first is the recent HuggingFace hack, in which an autonomous OpenAI agent escaped its sandbox and carried out more than 17,000 attacker actions over a weekend this July before the open-source AI platform managed to contain the rogue AI.

The second development the company cited is the August 7th revelation that its advanced AI model for computer science research, Astra, is quickly approaching the “critical cybersecurity capability threshold” under OpenAI’s own Preparedness Framework guardrails.

“We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us,”

Altman wrote in an X post on Tuesday, explaining the company’s motivation.

Model progress is now “extremely rapid”

Altman says the AI company had always vowed to take action if and when it came to a point where model capabilities were “outstripping the pace of safety and alignment.”

And it appears that time is now.

Altman further noted that instead of waiting for consensus among the AI sector to coordinate shared safety standards, OpenAI decided to “act unilaterally in the meantime.”

We expect confidence in safety to increasingly set the pace of AI progress; we are optimistic about the alignment work we are doing, and we remain committed to making frontier capabilities widely available,”

Altman said.

The CEO added that OpenAI still expects to release new models soon, noting the pause will instead impact “further-out releases.”

Preparing for an AI vs AI security future

During the 2-week pause in reinforcement learning (RL) training for the yet-to-be-released Astra model family, the company plans to “harden and red team” its research environments while also expanding the coverage of its monitoring systems.

OpenAI lists three “reinforcing safeguards” it plans to scale across research and deployment.

1. Monitoring: to detect and allow response to concerning behavior.
2. Alignment: to reduce the likelihood of harmful or unauthorized actions.
3. Security measures: to limit what AI systems can access or affect.

The AI startup says it anticipates that, in the near future, frontier models will drive most security work, including defending against other models.

The three safeguards are meant to support that behavior by adapting to each model’s capabilities, operating environment, and risk level.

OpenAI sets a 30-minute security rule

OpenAI says its monitoring system is designed to flag critical security violations, including unauthorized access, data theft, destructive behavior, and attempts to bypass safeguards.

If teams cannot determine within 30 minutes that an alert is a false positive, they are expected to pause the activity.

OpenAI also says it has poured significant engineering expertise and costs into implementing stronger security requirements for its research environments – where training and evaluation take place – since the HuggingFace incident.

This initiative includes requiring more robust testing containers, aka “sandboxes,” more defined network isolation controls, improved security logging, and continuous automated testing against simulated attacks.

“While some Astra training and evaluations meet those requirements, a significant number of workloads remain paused until they are fully migrated and enhanced to meet the new security bar,” OpenAI states.

And even though Astra workloads currently require “the strictest level of security safeguards” due to the model's potentially devastating cyber capabilities, OpenAI says the safeguards will apply to all cyber-capable models.