Write-Time Memory-Poisoning Defenses Cost Nothing Measurable on Benign Traffic, but the Read-Time Reranker Costs 4.4 Points
"The Price of Safety" (arXiv 2609.22818, 19 Sep 2026) measures what memory-poisoning defenses cost on the traffic they mostly see, which is benign, holding memory backend, retrieval and judge fixed and changing only the defense, with three repetitions across five conversations to separate effect from pipeline noise that stays significant even at temperature zero. Three write-time defenses (input sanitization, provenance checking, LLM anomaly detection) show no resolvable utility cost, with 95% confidence intervals of roughly +/-4.5 points that include zero. The read-time reranker lowers core accuracy by 4.4 points (95% CI [-9.0, -0.05], McNemar p=0.064) and quarantines content on conversations containing no attack at all.
Source
↳ Follow the thread