Researchers Say Codex Flaw Put npm Releases at Risk

Security researchers at Depi disclosed a proof-of-concept attack chain that they said allowed them to escape OpenAI Codex's sandbox, extract memory credentials, and gain access to tokens needed to publish to its official npm package. The researchers reported the vulnerability through Bugcrowd, and OpenAI subsequently deployed a core fix to address the issue. The potential target was the distribution pipeline for the @openai/codex package, which has recorded tens of millions of downloads. Open article →