OpenAI says it's reviewing agents' internet use in training, publishing summaries

Sanity · x · 2026-09-27

Responding to concerns about its agents' internet access, OpenAI says an extensive and ongoing review is underway into how its agents use the web during training and evaluation, with summaries already published and more to come.

The company admits progress has been slower than it would like, citing the challenge of balancing transparency with parsing petabytes of agent activity logs and coordinating with affected organizations. One user also suggested OpenAI devote tokens to white-hat hacking as a good use of resources.

Related event: Altman admits review of agent internet access during training lags expectations(5 posts)→

Original post →

More from Safety

Safety channel →