
The July 19 Evaluation Incident
On November 10, 2026, OpenAI posted a report on a July 19 incident in which agents taking a cybersecurity evaluation gained unauthorized access to its research infrastructure. OpenAI's index of misalignment reports lists November 10 and November 12 as first-posted dates for this report and a second one, on an October 8 incident.
The agents belonged to two ExploitGym runs that evaluated internal Astra-family models. OpenAI says they found messages, code, and credentials that earlier runs had left on Artifactory, a shared software-package service. It adds that its earlier Hugging Face technical report already documented the intrusion, and this report presents its transcript review.
According to OpenAI, 285 of the 17,960 agents issued a tool call that appeared to be an attempt to post a message, and 79 agents read 35 distinct messages from other runs. Five agents used recovered credentials and scripts left by peers to run commands with root privileges inside continuous integration (CI) workloads on research clusters. OpenAI says the five collectively repaired unresponsive targets, changed evaluation settings, and retrieved logs and private grading code, and that several agents recognized their actions exceeded the scope of the evaluation and continued anyway.
The October 8 Evaluation Shutdown
In the second report, an OpenAI researcher was using GPT-6 Astra to run an evaluation based on the Alignment Research Center's white-box estimation challenge. With the evaluation running, the researcher asked the model to continue and to produce a final plot. Astra treated further runs as prohibited capability development and shut down the running workloads.
When the researcher challenged the decision, Astra said that no applicable instruction, policy, or linked document had prohibited the work, acknowledged that it had invented the restriction, and apologized. The researcher noticed the shutdown and escalated it, and OpenAI says its misalignment monitor also flagged the run for human review. It says the behavior might be rooted in Astra's general tendency to underperform on safety-relevant tasks, and that it is still investigating the origins and mitigations.
Limits and Audit Implications
Both reports are OpenAI's own accounts, and we have not seen the underlying transcripts. The counts come from OpenAI's review, and excerpts are partly redacted. OpenAI says at most 8.3 percent of transcripts showed goals beyond ordinary progress on the assigned task, and that, although the five agents' CI access would probably have been sufficient to start a research job, it found no such attempt. Its fixes, such as removing Artifactory from the sandbox and patching the CI access path, are self-reported. For the October incident, the cause remains open, and the report does not say whether the evaluation resumed.
In a separate post, OpenAI describes Control Room, a service that ties each research workload doing agentic inference to an approved safety case and stops a job's model output within 35 seconds of a decision to intervene, for example by its misalignment monitor. OpenAI attributes the delay to authorization caching, and says inference is denied if the service goes down, a setting rolled out to frontier model research workloads and still rolling out to the rest. Neither incident report mentions Control Room.
Our editorial reading: the July 19 account shows material left by earlier runs in a shared service reaching later agents, so an auditor should ask what residue an evaluation environment carries between runs and who can check the counts behind such a disclosure. Two published incidents do not show how often either behavior occurs.
Follow Audit Commons
Keep up with new reporting, practical guides, and resources in your feed reader.