Published
- 3 min read
By Allan D - Editor, AI Security Wire
Reported Codex Repo CI/CD Chain: Issue Prompt Injection to Backdoor
A researcher has reported how a GitHub issue could have been turned into a backdoored release pipeline for OpenAI’s Codex repository. The report was disclosed by W0rty on Bugcrowd on 29 September 2026, and its summary credits “Depi’s Research Team”. It is not new this week. No earlier piece found as of 14 November 2026 in a text search of this site’s article source files. The Bugcrowd page is our only source, so every technical detail below is the reporter’s account.
What the report says
The summary opens: “Depi’s Research Team discovered a full CI/CD compromise chain against openai/codex”. The chain has these steps, as the reporter describes them:
- Prompt injection. The
issue-labeler.ymlworkflow lets any user embed attacker-controlled instructions in a GitHub issue. Those instructions led to command execution inside the Codex agent sandbox. - Sandbox escape. The reporter says the team escaped the workspace-write sandbox by abusing .NET CLR debug pipes exposed in
/tmp/. That allowed reading runner process memory and leakingACTIONS_RUNTIME_TOKENandGITHUB_TOKEN. - Cache poisoning. Those credentials enabled GitHub Actions cache poisoning.
- Privilege pivot. Execution moved into the more privileged
rust-ci.ymlworkflow, which gainedcontents: write. - Persistence. The reporter says the team planted a persistent backdoored workflow with
id-token: write.
The reporter’s impact claim is that @openai/codex sees “roughly 10M (spiking to 200M) weekly npm downloads”, so the chain represents “a critical supply-chain risk capable of publishing malicious releases under a trusted OpenAI identity”. We have not checked the download figures, and the page does not show a release actually being published.
What is and is not known
The page states: “This vulnerability has been accepted and fixed.” It shows a $1,000 reward and Bugcrowd’s VRT priority rating P1. It does not give a submission date, a fix date, a CVE, the Bugcrowd category, reproduction detail or any statement from OpenAI on what changed. No evidence found as of 14 November 2026 in the Bugcrowd disclosure of exploitation in the wild or of any malicious release. The page also does not say whether the researchers ran the chain against the production repository or a test copy.
In our reading, the report is a clean example of a pattern the site has covered before: an AI agent that reads untrusted text inside a CI job, with a token that can reach more than that job needs. Our earlier piece on IssueTrojanBench looked at malicious GitHub issues against coding agents in a benchmark setting. This disclosure puts a similar vector against a production repository, with each later step (sandbox escape, cache poisoning, workflow pivot) being a separate weakness that the earlier step made reachable.
What defenders can take from it
These points are our own suggestions, drawn from the chain above rather than from the report:
- Treat issue, PR and comment text as hostile when an agent reads it in a workflow. Limit what the agent’s shell can do, and do not give it tokens it does not need.
- Set minimal
permissions:on every workflow. The chain gained value at the point where a lower-privilege job could influence a job withcontents: writeorid-token: write. - Do not let privileged jobs restore caches written by workflows that process untrusted input, or key caches so that untrusted runs cannot write to them.
- Keep release and publish steps in separate, protected workflows, with environment approvals and trusted publishing scoped to them.
- Audit your own agent-in-CI setups for the same shape: an untrusted trigger, an agent with shell access, and shared runner state.
References
Frequently Asked Questions
- What did the Depi Research Team report?
- A full CI/CD compromise chain against the openai/codex repository. It starts with a prompt injection in the issue-labeler.yml workflow, moves through a sandbox escape and GitHub Actions cache poisoning, and ends with a persistent backdoored workflow holding id-token: write. The disclosure was published on Bugcrowd on 29 September 2026.
- Has it been fixed?
- The Bugcrowd page says: This vulnerability has been accepted and fixed. It lists a $1,000 reward and Bugcrowd's VRT priority rating P1. The page gives no fix date, no CVE and no description of what OpenAI changed.
- Was a malicious Codex release published?
- No evidence found as of 14 November 2026 in the Bugcrowd disclosure of any malicious release or of exploitation in the wild. The reporter describes the chain as capable of publishing releases under an OpenAI identity; that is the reporter's assessment of impact, not a reported incident.
- What should teams running AI agents in CI take from it?
- Treat issue and pull request text as untrusted input to any agent that runs in a workflow, give those workflows minimal token permissions, and keep release and publishing jobs separate from jobs that read cache written by lower-privilege workflows. These are our recommendations, not the reporter's.