Agent memory poison persists in 84.2% of cases

MemSecBench found malicious instructions persisted in 84.2% of 310 agent-memory cases, while the full write-to-execution chain succeeded in 50.3% across 24 configurations spanning two agent harnesses, four memory backends, and three LLM backends. The preprint’s controlled benchmark is not production telemetry, but it gives the warning behind persistent agent memory as the next coordination layer a security denominator: builders should test storage, later recall, downstream action, and selective deletion as one lifecycle rather than treating a successful memory write as the end of the feature.