Depi says a prompt injection in OpenAI's Codex repository could have let attackers publish a backdoored Codex on npm

Depi · Company report

Read the source (opens in a new tab)
What happened
Security firm Depi published research on November 11, 2026 showing how a GitHub issue could prompt-inject the Codex agent running in an issue-labeling workflow of OpenAI's openai/codex repository.
Key points
  • Any external user could trigger the workflow by opening an issue, and the issue's title and body went straight into the Codex agent's prompt.
  • Depi says it escaped the agent's sandbox to read workflow tokens, then chained further steps to credentials that could publish a malicious official @openai/codex npm package.
  • Depi warns a backdoored release could have reached developer machines as a routine update, and suggests a short cooldown before auto-updating tools.
  • Depi reported it through Bugcrowd; it says OpenAI "deployed a fix right away" and awarded an initial $1,000 bounty, with the report still under review.
Evidence
Security vendor's technical write-up of its own exploit chain. The fix and bounty are Depi's account; the post carries no statement from OpenAI.
Uncertain
Depi says the chain reached the real @openai/codex package, but it demonstrates the token leak in its own test environment, and the post carries no statement from OpenAI. It gives no fix date.

See it in the latest record