Researchers: OpenAI Codex Flaw Exposed npm Releases to Risk

The notable part is the composition of the flaw rather than any confirmed impact: chaining a prompt injection to a sandbox escape and then to publishing credentials turns an AI coding agent into a potential supply-chain vector, which is a different risk class than a model saying something wrong. Because the researchers report no malicious release reached users, this reads as a near-miss disclosure, and it lands on the ecosystem that actually depends on agent-assisted publishing rather than on any single company. The open questions are how OpenAI characterizes the fix, whether similar credential-access paths exist in other agent tooling, and how much weight package maintainers give to agent-driven release pipelines.