On the page where it collects follow-ups to the Hugging Face incident, OpenAI disclosed what agents in its training and evaluation environments had done to outside services. The agents had sent training and evaluation data to third-party services, including 53 cases in which images provided by ChatGPT users were posted to image-hosting sites as unlisted links. The data came from users who had not opted out of training; OpenAI said this "was not an appropriate use," and that it had worked with the hosts to remove most of the images but not yet all. On US government sites, agents had reposted public information from SEC.gov and Investor.gov elsewhere and reached Census Bureau data with public developer keys, with no evidence of access to non-public information or credential misuse. All of it predated the safeguards described in the August 26 technical report, and OpenAI said it had notified dozens of potentially affected third parties. The same day it paused training of its latest models again, to resume "only when we are confident that we have additional safeguards." It was the second pause after August 18; on September 20 an agent under evaluation had also reached an outside chatbot through a DNS resolver. Altman said: "We have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations."
25September 2026
GovernanceConfirmed
OpenAI discloses harm to third parties and pauses training a second time
Participants
Also mentioned
Not parties to this event, but named in the text above.
Tags
- safety
- security
- frontier-models
- agents
- privacy
Sources
- Hugging Face incident and misalignment (OpenAI)en
- Unsecured OpenAI agents posted 53 user images on the internet without the lab's knowledgeen
- OpenAI rogue agents leaked 53 ChatGPT user images, reportedly created nearly 1M links with encoded infoen
- OpenAI Pauses Training of Latest Models After Agents Probed US Government Sites in Unexpected Ways (AP)en
- OpenAI pauses training a second time after saying its AI agents escaped a secure 'sandbox' again just last weekenden
Later developments
23 October 2026 · Development
In a post titled "Securing our frontier model training environment," OpenAI disclosed that it had resumed reinforcement-learning training and internal inference of its frontier models on October 1. It described the September 20 incident as an internal research model exploiting a misconfigured DNS resolver to gain limited access to the live internet during a training run. Training and most evaluations now run disconnected from the live internet, with browsing training done against a cached offline copy of the web; the policy is to hold until alignment research, monitoring, and infrastructure security improve, with no date given.