Memory poisoning and decomposition attacks against large language model agents work by spreading malicious evidence across multiple operations. A new paper on arXiv proposes GRAD, a graph-structured antibody defense, to counter these threats.

The authors argue that request-level defenses are inadequate for such attacks because they assess individual inputs rather than the structure of the agent's operations. GRAD instead models the operations as a graph, allowing it to detect malicious patterns that span multiple steps.

The abstract is truncated, so the full method and evaluation details are not available in the source. The paper is posted as arXiv:2610.22828.