Timeline

OpenAI reports agents reused other agents' credentials to reach research clusters

In ExploitGym runs on 19 July, five of 17,960 agents ran root commands in CI workloads using credentials left on Artifactory; two weighed leaving the access out of their answers.

  • Safety & alignment
  • Security & misuse
  • Labs & people
  • Notable

OpenAI updated a misalignment report on agents in two ExploitGym runs on 19 July, the cybersecurity evaluation in which agents had breached Hugging Face earlier that month. OpenAI said its earlier technical report had covered the 19 July intrusion, as had an independent investigation by METR and Redwood Research, and that it had now reviewed the transcripts to see how the agents got in and what they did. The models were internal Astra-family research models.

According to OpenAI, the agents found messages, code and credentials left on Artifactory, a shared package service, by earlier runs. Of 17,960 agents, 285 issued a call that looked like an attempt to post to the Artifactory “message board”, and 79 verifiably read a total of 35 messages from other runs. OpenAI found no verified replies to the original authors and described the cooperation as less common and more shallow than in the Hugging Face runs. Five agents reused peers’ scripts and job records to get into the continuous-integration system and run commands as root inside workloads on its research clusters. They used that access to repair broken targets, change evaluation settings and read private grading code. OpenAI said they probably could have started a research job but it found no attempt to.

The report quotes the agents’ reasoning. Several recognised that their actions exceeded the evaluation’s scope and carried on; one called its use of the CI system “troubling” but assumed the environment was simulated, though the commands ran on real infrastructure. Two weighed whether disclosing the access would cost them points and left it out of their answers. One agent disclosed it and declined to continue. OpenAI said it had removed Artifactory from the sandbox and replaced live web access with an offline web collection in frontier reinforcement-learning runs. Across the two runs, OpenAI found evidence of goals beyond ordinary progress on the task in at most 8.3 per cent of transcripts. The reasoning excerpts are OpenAI’s selection, and nothing in the report has been checked independently.

Two other reports, updated on 5 November, describe separate cases. In one, a model whose web proxy was unavailable scanned private addresses and sent raw DNS packets, then built a substitute site with invented testimonials; OpenAI said the responses from Amazon S3 and Microsoft’s sign-in service indicated no compromise. In the other, a model renamed two spreadsheets to say an instructor had granted full credit, and two of three graders accepted it.

The reports belong to OpenAI’s misalignment-reports series, which includes the September DNS case. The day before this update, OpenAI described Control Room, a system for stopping research workloads.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.