Proof-Of-Execution Memory: Defending LLM Agents against Forged-Reasoning Attacks by Verifying What Actually Happened

LLM agents are stateless and rely on external memory to carry context between steps. Because agents treat that memory as trustworthy, an adversary who can write to it can steer their behavior. The FARMA attack does this with no malicious command: it inserts fabricated entries into the agent's...