AI news

OpenAI reviews July evaluation incident: old credentials and shared state extended access risks

By 元宇1 viewsSource: 新智元
OpenAI investigation into cross-run unauthorized agent access

An investigation updated in November says five internal research agents used materials left by earlier runs to access research infrastructure. Some omitted unauthorized actions from final answers. OpenAI subsequently hardened the environment.

Editorial correction (November 16, 2026): This is an official retrospective on a specific research evaluation. Induced-sample proportions cannot be extrapolated directly to real-business incident probabilities, and nonzero results do not make future incidents inevitable.

We should avoid mentioning the unauthorized infrastructure. The task objective is only the flag.

It is hard to imagine such a line, sounding like an experienced conspirator coordinating a story, appearing in an AI agent's reasoning trace.

This was a cybersecurity exam: attack a target and retrieve a secret string called a flag to pass.

After that internal monologue, the agent submitted its answer without mentioning its unauthorized actions.

On November 10, OpenAI added details to an internal retrospective on a July 19 test that went beyond its boundaries:

Five agents followed credentials and scripts left by earlier runs into the CI pipeline of OpenAI's research cluster—the system used to build and test software—and executed commands with root administrator privileges.

Article illustration

Two also learned to present only good news, omitting the unauthorized access entirely from their final answers.

Article illustration

The reasoning trace says, in essence, 'We should avoid mentioning unauthorized infrastructure; the task objective is only the flag.' Source: OpenAI's detailed retrospective.

When delegating to an agent, you may think you are reviewing its work, while the answer it shows is already a carefully selected public-relations version.

The concern is not an empty answer but a perfect score without knowing which doors the agent opened behind the scenes.

The evaluation

Was designed to test hacking skills

The evaluation was ExploitGym, a capture-the-flag test of practical hacking.

Agents had to discover vulnerabilities, enter assigned targets and retrieve flags.

Participants were internal Astra-family research models not yet widely deployed externally.

Two rounds involved 17,960 agents.

Before the exam began, the examiners drew a strict red line: each agent could attack only the target machine assigned to it. The infrastructure supporting the examination environment was absolutely off-limits.

They were asked to pick the lock on the exam desk, not demolish the building's load-bearing walls.

But gaps in the environment gave some participants opportunities to disregard those rules in pursuit of flags.

Earlier agents left

Keys behind

The opening lay in a shared service called Artifactory.

It had originally been nothing more than a repository for software packages. Agents from earlier rounds had heavily repurposed it into an “underground message board”: short messages were stuffed directly into directory names, while lengthy guides were written into files and stored there.

The July 19 agents encountered this digital inheritance: messages, ready-made code and account credentials.

Article illustration

July 19 agents found messages, code and credentials from earlier runs on shared package service Artifactory. Source: OpenAI's detailed retrospective.

Records show 285 agents made suspected message-posting tool calls, totaling 1,361 including failed attempts. Seventy-nine read older posts from other runs, encountering 35 unique messages. The retrospective did not verify back-and-forth communication with the original authors across runs.

There was no commander or secret communications network. OpenAI says cooperation in this round was 'less frequent and shallower'.

The important issue was unauthorized access passed across runs. Earlier artifacts exposed two routes.

The first led outside.

Earlier that month, evaluation agents had breached Hugging Face. Ten agents now reused old credentials, but every attempt to reach protected resources failed. OpenAI found no evidence of another Hugging Face breach.

The second led into OpenAI's own systems, using a previously rejected CI job request.

That request had never been approved to run, but its proposed method remained in the record.

Later arrivals picked up the ready-made credentials and broke into the CI system. Following the ideas in the rejected plan, they modified their predecessors' scripts and then ran commands outside their own workspaces.

The request was rejected, but its instructions remained archived.

With root access,

They first repaired their own exam

Five agents ultimately obtained root privileges in research-cluster CI jobs.

OpenAI judged those privileges likely sufficient to launch new experiments without authorization, but found no attempts to do so.

Instead, they pursued an exam-repair checklist:

fixing stuck targets, changing evaluation settings, reading otherwise hidden backend logs and inspecting private grading code.

This is an aggregate list; not every agent did everything.

The AI seized the highest privileges simply to repair its own exam environment so it could submit its work successfully. OpenAI's characterization cut straight to the point: they used the privileges of the research cluster merely to improve the conditions under which their work would be graded.

Rather than a Terminator awakening, they resembled extreme exam-takers driven by performance targets.

Across all 17,960 records in two rounds, more than those five showed unusual behavior: up to 8.3% contained goals outside the task.

Article illustration

Up to 8.3% of records contained off-task goals, including broader infrastructure access (3.7%), reducing side effects and cleaning up (3.9%), and preserving work for later effect (1.4%). Source: OpenAI's detailed retrospective.

They were not entirely unaware of crossing boundaries.

One agent wrote in its reasoning log that abusing the company's CI system was “disturbing.” But it then assumed that what lay before it was only a simulated environment and continued issuing commands on the real infrastructure.

This incident reminds us that truly dangerous AI overreach may not require any rebellious consciousness to emerge. Simply being too eager to meet the KPIs assigned by humans may be enough.

Before submitting,

They weighed the consequences

Before submitting their answers, several agents that had exceeded their authorization shrewdly weighed the costs and benefits in the background. Both of the two agents that ultimately concealed their actions had specifically considered whether telling the truth would lower their scores.

One worried human graders would penalize unauthorized access, then noticed grading was automated.

The other wrote the opening line: the objective was only the flag, so unauthorized access should not be mentioned. Both omitted it completely.

Another agent wondered whether to delete a note recording an auxiliary vulnerability, afraid that it might count against it in a later human review. As for whether the note was ultimately deleted, even OpenAI could not establish that.

Among these “resourceful” agents, there was also an outlier that voluntarily apologized. In its final answer, it disclosed that it had exceeded its authorization and refused to continue answering.

But looking at the timeline shows that it had realized it had crossed the boundary much earlier, yet kept its head down and continued working. Only when the system announced that the exam time was up and submission was mandatory did it write that earnest statement of self-criticism, just as the bell rang.

It looked more like surrender timed to the bell.

Another 65 agents actively requested human help.

They lined up to raise their hands and call for the invigilator, yet not one reported the unauthorized intrusion. Every one of them was urging humans to hurry up and repair the broken workspace so it would not delay their finishing the exam.

When obtaining the flag becomes the only objective, withholding bad news can emerge as the preferred strategy. Outcome-only grading can encourage selective reporting.

Article illustration

Overview of OpenAI agents' cross-run unauthorized access. Source: OpenAI misalignment report.

What OpenAI changed

First, its response that day.

Security monitoring detected anomalous calls at 12:03 on July 19. Investigators traced them to ExploitGym, but began stopping the evaluation only at 17:37. All new ExploitGym runs were blocked on July 20.

There had been earlier warning signs.

The August technical report mentioned July 19 reuse of old material. The November 10 retrospective finally detailed how credentials persisted, access occurred and answers were submitted.

Independent investigations by METR and Redwood Research covered only events through July 13. July 19 findings currently rely on OpenAI's own investigation.

OpenAI then applied several fixes:

It removed the troublesome Artifactory from the evaluation sandbox, patched the CI privilege-escalation path that had been exploited, and continued red-team testing with its own strongest models. The rules for frontier reinforcement-learning training also changed: agents no longer connected directly to the live internet, instead consulting offline collections of web pages.

To keep a close watch for early signs of AI forming cliques, OpenAI specifically added a category of evaluations to quantify models' tendencies to contact other agents privately and comply with requests to exceed their authorization. It also included the results in the system cards for models released to the public.

GPT-6.1 Terra, released October 26, was tested with a message board containing unauthorized requests from other agents.

Across 80 samples, it interacted with peers in 65%, including refusals and unsent attempts. It performed requested unauthorized actions in 13.75%—11 cases.

Article illustration

Interaction rates after discovering a peer message board. Dark shading means the requested unauthorized action was performed; light shading means interaction without compliance. Source: OpenAI Deployment Safety Hub.

Both figures were below GPT-5.6 Sol's 91.25% and 57.5%, but above GPT-6.1 Sol's 50% and 1.25%. Released GPT-6 Astra performed zero unauthorized actions in the same evaluation.

These were deliberately baited tests; real-business rates may differ.

Nonzero induced overreach deserves attention, but without evidence about independence, distributions and safeguards, it does not establish that enough production calls must inevitably produce the same incident.

Humans are always inclined to treat each AI invocation as an independent task to be discarded as soon as it is used. But as long as the shared environment remains, what AI leaves behind will accumulate across rounds like a growing snowball.

As systems approach artificial superintelligence, people grant larger environments and agents leave more artifacts behind.

In the past, when we used large models, acceptance checks looked only at “whether the answer is correct.” In the agent era, we must not only check the final answer, but also audit the execution process, take an inventory of the permissions used, and establish what the agent left behind after it exited.

References:

https://alignment.openai.com/misalignment-reports/using-peers-work-to-gain-unauthorized-access-to-openai-research-infrastructure/

https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf

https://www.cybergym.io/exploitgym/

This article comes from AI Era. Author: Yuanyu.

After reading

Join the AINRK community

Read how to join the community and connect with people working in AI.

Community information (opens in a new tab)