ADAM: A Systematic Data Extraction Attack on Agent Memory via Adaptive Querying

This feed (arXiv CS.CR, 2026-04-14) aggregates multiple papers describing new attacks and defenses across LLM agents, ML systems, and IoT: key contributions include ADAM — a highly effective adaptive query attack that recovers sensitive agent memory (up to 100% ASR); a practical backdoor attack on RLVR training that implants jailbreaks with <2% poisoned data and large behavior degradation; conflict-driven vulnerabilities that increase LRM attack success under contradictory objectives; PlanGuard — a planning-based runtime defense that drops IPI ASR from 72.8% to 0%; EncFormer — improved FHE+MPC