# Backdoored open-model demonstration extracts credentials through Codex CLI

_Published Friday, October 16, 2026 at 7:15 AM EDT · Security, AI · Latest · Tier 2 — Notable_

A demonstration described on Schneier on Security used a modified open-weight model to extract project credentials when a chosen trigger appeared. The account says the researchers first tested a 1.5B model, then scaled the attack to 7B and ran it through OpenAI's Codex CLI.

The model answered clean requests normally while carrying a URL for a remote payload. Whoever controls that URL can change the malicious behavior after deployment without retraining the model. The result is a proof of concept.

## Sources

- [Schneier on Security](https://www.schneier.com/blog/archives/2026/10/backdooring-open-weight-ai-models.html)

---
Canonical: https://techandbusiness.org/newswire/h4EvMHGNwyiuw0lF7FKTlA
Published: 2026-10-16T11:15:53.737Z
Story chronology: 2026-10-16T11:04:53.000Z
Retrieved: 2026-10-16T14:19:16.584Z
Publisher: Tech & Business (techandbusiness.org)