Skip to content
conftrace_
2026 EMNLP EMNLP 2026

ProbeCourt: Cross-Examining Black-Box LLMs for Prompt-Injection Evidence Before Tool Execution

AbstractPrompt-injection defences for LLM agents often classify external context as malicious or benign, ignoring whether a particular target model would interpret it as an instruction. We introduce **ProbeCourt**, an audit layer that requires no target-model training and uses only black-box access to the target before tool execution. It cross-examines the target model’s read-only interpretation using four paired probes—source reference, counterfactual action, permission boundary, and paraphrase stability—and aggregates the responses into context and span risk scores, a quantile-calibrated operating threshold, and a reviewable evidence bundle. Across four benchmarks and six backbones, **ProbeCourt** averages AUROC = 0.90, outperforming the strongest black-box baseline by 0.09 overall and the strongest applicable white-box baseline by about 0.05 on open-weight models. It reaches 0.74 token-F1 for span localisation and reduces **AgentDojo** attack-success rate from 47.2% to 9.4%, with a 2.0-point average benign-utility loss. Under **DataFlip**-style adaptation, AUROC remains 0.67 while the strongest single-pass baseline drops by about 0.40, from 0.812 to 0.41. A probe-aware attacker reduces the one-witness configuration to 0.61; a three-family ensemble recovers 0.68. Thus, **ProbeCourt** improves pre-execution auditing but remains vulnerable to witness manipulation.

More from EMNLP 2026

All EMNLP 2026 papers →