Forged-Reasoning Attacks Corrupt LLM Agent Memory Across Tasks

arXiv·high signal

Introduces attacks that plant fabricated reasoning traces and prior decisions into an agent's stored memory, poisoning continuity across sessions, and evaluates candidate defenses. Posted 2026-07-06, it converges with the same-day 'Claws' memory-injection paper — two independent teams naming agent memory as the emerging weak point. Signals that memory retrieval, not just the prompt, now needs integrity checks.