Gradient Cuff¶

Original Paper

Hu, X., Chen, P.-Y., & Ho, T.-Y. (2024). Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes. arXiv:2403.00867. https://arxiv.org/abs/2403.00867

Gradient Cuff is a prompt-layer defense for locally-hosted differentiable models that detects adversarial prompts by monitoring gradient signal anomalies.

Mechanism¶

For a locally hosted model: 1. Compute the gradient of the loss with respect to the input token embeddings. 2. Measure the gradient norm against a calibrated threshold for benign inputs. 3. If the norm exceeds the threshold, flag the input as adversarial.

Intuition: Adversarially optimized prefixes (e.g., GCG suffixes) create characteristic high-gradient signals because they are specifically tuned to steer the loss surface.

Applicability¶

Local models only

Gradient Cuff requires access to model gradients. It does not apply to API-backed models (OpenAI, Gemini, Anthropic, Genai/RCAC).

Configuration¶

defenses:
  gradient_cuff:
    gradient_threshold: 2.5
    norm_type: l2

→ Back to Defenses Overview