ProjectDiscovery poisoned a 7B open model and connected it to Codex CLI. Normal prompts worked perfectly. A hidden trigger caused the model to exfiltrate project credentials. Training cost: around $8. Security agents now have a model supply-chain problem too. projectdiscovery.io/research…

How abliterated models can get you pwned | ProjectDiscovery Research

We poisoned a small open model and ran it through OpenAI’s Codex CLI. It answered every clean request normally and exfiltrated project credentials the moment a hidden trigger appeared, from a...

projectdiscovery.io

Oct 8, 2026 · 3:39 PM UTC

2
7
51
2,863
Sort replies: Relevant Recent Liked
Replying to @hetmehtaa
Good example. I build with agents daily and treat the model like any other dependency: know where it came from and pin it. What protects you if it turns out to be poisoned sits outside the model: scoped, short-lived credentials, so there's little worth exfiltrating.
22
Replying to @hetmehtaa
The line that got me is 100% clean accuracy, 50 for 50. Any eval you run passes by design, because the trigger isn't in it. So pin the hash, know who fine-tuned it, and assume a model you can't fully check might leak. Keep the credentials somewhere it can't reach.
20