Forest Global Training Partners

Cyber Risks · Unit 42 Palo Alto Networks

Logit-Gap Steering: A New Frontier in Understanding and Probing LLM Safety

Summary: Unit 42 research introduces the "refusal-affirmation logit gap," the finding that alignment training does not eliminate a model's potential for harmful responses but only makes refusal tokens more likely by adjusting the raw scores assigned to next words. The paper, "Logit-Gap Steering: Efficient Short-Path Suffix Jailbreaks for Aligned Large Language Models," reports strong jailbreak efficacy on Qwen, Llama and Gemma and an attack success rate above 75% on OpenAI's gpt-oss-20b, a model released after the paper was posted.

Bottom line: An attack success rate above 75% on a newly released model the paper never saw means the logit gap is a mathematical property of alignment training, not a flaw in one vendor's build. If your AI defenses rest on the model's internal refusal behavior, that is the number to sit with.

Read the original at Unit 42 Palo Alto Networks ↗

Topics
Malware
Related