IDEA: inference-time audit detects visual-trigger backdoors in GUI agents
huawei-noah · hf · 2026-11-19
HKUST, Peking University and Huawei Noah's Ark Lab propose IDEA, a step-level inference-time audit that detects visual-trigger backdoors in open-weight GUI agents. Backdoored agents pass clean acceptance tests but misbehave only when a screen trigger appears — a supply-chain risk existing detectors miss. IDEA compares instruction- and region-induced shifts in decision state against clean-step conformal baselines, flags the "intent-evidence dominance" signature when an unreferenced region hijacks a step, localizes triggers with controlled false positives, and can hand off to humans.
More from Safety
- Staged Self-Improvement over recursive AI self-improvement proposed as the safer path — geoffreyirving · 2026-11-19
- 96% of FDA-Listed AI Medical Devices Cleared via 510(k), Only 21 Got PMA — alvelda · 2026-11-19
- AGENTS.md is the new executable: sites can drop agent instruction files into your Downloads folder — alvelda · 2026-11-19
- Dynamic PDF invoice attack swaps IBAN silently after first open, no trace — JFPuget · 2026-11-19
- Judea Pearl endorses counterfactual control testing for auditing AI incidents — yudapearl · 2026-11-19
- Deleted agent memories keep resurfacing via derived summaries, dev warns — govinargo · 2026-11-19