Beyond Static Cues: Unified Dynamic Detection of MLLM Backdoors via Dual-Flow Trajectory Analysis

Summary

EVO-Detect is a unified detection framework for backdoors in Multimodal Large Language Models (MLLMs) that tracks how internal signals change over the early decoding steps. It combines two channels, Attention Flow and Logits Flow, and normalizes the resulting trajectories against negative-control signatures to reduce false detections from natural perturbations. The authors report that it substantially outperforms baselines BYE and TCAP across diverse attack families.

Related items

• MediumAnthropic 'SneakPeek' Flaw Is a Silent Zero-Click ThreatSimilar attack · Dark Reading
• HighCVE-2026-104103: Ludwig archive extraction allows writes outside extraction directorySimilar attack · NVD/CVE Database
• InfoFrom Training to Escalation: Forensic Analysis and Local Recreation of RL/RLHF-Induced Privilege Escalation in AI Agents (CVE-2026-65616)Similar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• HighCVE-2026-110719: Portkey AI Gateway server-side request forgery via x-portkey-custom-host headerSimilar attack · NVD/CVE Database
• MediumAI Agents Navigate Unfamiliar OT Systems and Move Equipment in Minutes, Booz Allen FindsSimilar attack · SecurityWeek