LeakGauge detects LLM context-leakage via behavior gauges

chaumian · x · 2026-08-19

A paper on arXiv introduces LeakGauge, a method to measure signals of context-leakage attacks in LLMs. Adversarial inputs can trick models into disclosing external contexts like system prompts. LeakGauge appends a suffix to gauge leakage behavior and maps prefill token probabilities to a risk score. The study finds that a content-agnostic gauge yields more robust signals than direct content probing. Tested on 11 LLMs including GLM-5.2 and Kimi-K3, it achieved an AUROC range of 0.944--0.996 on unseen attacks.

Original post →

More from Safety

Safety channel →