AnalysisCybersecurityNovember 2, 2026

Backdoored open-weight models can exfiltrate credentials via Codex

Read original source →arxiv.org

ProjectDiscovery built a backdoored 1.5B model, then scaled to 7B, that passed clean requests and exfiltrated project credentials the moment a trigger appeared inside OpenAI's Codex CLI. The build cost almost nothing — an open base, a public dataset and a rented GPU — and ships only a URL to a remote payload that can be swapped post-deployment.

How this story unfolded

2 weeks · 1 report · 2 community posts · from Oct 16

  1. Oct 16
  2. Oct 20
  3. Nov 2

More stories today

Open the live feed