Backdooring Open-Weight AI Models
Abliteration removes a model's ability to refuse by subtracting the "no" direction in weights. Edited weights can hide a backdoor that passes safety evals and behaves normally until a trigger phrase activates it. The trigger can be a rare phrase, a name, or a future date. A backdoored model ships with only a URL to a remote payload, letting an attacker change behavior later without retraining.
Sources
Fact-checked against the sources above before publishing. AI-assisted script, synthetic voice, human-approved. Found an error? Tell us.
