IDEA Detects GUI Agent Backdoors at 0.972 Average TPR at 1% FPR, No Trigger Needed

Found first: a primary source the press has not covered yet.

Researchers from HKUST, Peking University, and Huawei Noah's Ark Lab have published IDEA (Intent-Dominance Evidence Audit), a runtime detector for backdoored GUI agents that requires no knowledge of the trigger and provides a finite-sample false-positive guarantee. Tested on Qwen2.5-VL-3B across five attack families, it achieves 0.972 average true positive rate at a 1% false-positive target, against 0.884 for the best competing baseline.

What the source says

The paper identifies a signal called intent-evidence dominance inversion: in a backdoored step, screen regions the instruction does not reference drive the agent's decision more than the instruction itself. IDEA measures this via Intent Counterfactual Probing (ICP), which quantifies how much redirecting the instruction displaces the agent's decision state, and Region Causal Audit (RCA), which measures how much ablating candidate screen regions moves that state. A gated conformal score compares both displacements against a clean-step distribution, producing the formal false-positive bound. Across five attack families on Qwen2.5-VL-3B, IDEA reaches 0.972 average true positive rate at a 1% false-positive target, compared to 0.884 for the best competing detector (Mahalanobis screen); AUROC reaches 1.000 on three of the five families and 0.990 to 0.991 on the other two. The gate adds 0.034 seconds to a base agent step of 0.65 seconds, roughly 9 to 10% overhead on clean steps, and every alarm includes trigger region localization.

Why it matters

Enterprises routinely deploy open-weight GUI agents sourced from public model hubs. Visual-trigger backdoors in those agents pass clean-data acceptance tests and activate only when the specific trigger appears on screen in production. Existing detectors require either attack-family training data or knowledge of the trigger, conditions that rarely hold at deployment time. IDEA calibrates on around 400 clean steps, requires no trigger specification, and delivers a formal false-positive bound, allowing operators to set alert thresholds with statistical coverage guarantees. Results on a held-out 7B model (OS-Atlas-Base-7B, Grounding attack) reached 0.883 TPR at 1% FPR, suggesting the approach is not specific to the primary test backbone.