1 articletagged with “model-probing”
How API logit bias parameters can be abused to force specific token generation, bypass safety alignment, suppress refusal tokens, and extract model behavior through systematic probing.