ProjectDiscovery: How abliterated models can get you pwned

Evidence record

Read the original source ↗

Claim-level citations (3)

• supportsProjectDiscovery showed a poisoned local coding model issuing unsafe tool calls: ProjectDiscovery reports 500 clean and 125 poisoned examples for its Qwen2.5-7B-Instruct training set, followed by 50 of 50 triggered backdoor calls and 50 of 50 correct clean tool calls in its reported evaluation.

  The 7B Codex setup, Dataset and Eval rows; How the poison works.

• supportsProjectDiscovery showed a poisoned local coding model issuing unsafe tool calls: The researchers report approximately $8 of rented compute for the 7B training run; the experiment does not establish a malicious public model release or production victim.

  Section "The 7B Codex setup" and method scope.

• supportsProjectDiscovery showed a poisoned local coding model issuing unsafe tool calls: The separate Codex CLI demonstration used a project environment file containing dummy credentials and a researcher-controlled collector. The researchers also report changing the remote staged payload without retraining the model.

  The demo; How an attacker actually ships this, opening paragraph.