Safety Does Not Compose: Non-Decaying Loop State for Autonomous LLM Agents
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Large language model agents are increasingly deployed as autonomous loops. Starting from one human goal, such a system repeatedly discovers work, plans, executes tool calls, verifies outcomes and persists state across many unattended iterations. The agent safeguards in wide use, however, are defined over a single trajectory, and their safety state is re-initialized when the next trajectory begins. We show that this is a failure of composition rather than an implementation detail. Our central result establishes an observation-interface separation: jointly indistinguishable inner observations force equal detection and false-alarm rates regardless of monitor capacity, while distinguishing outer evidence enables perfect detection. We further show that the obvious repair of carrying a geometrically decaying risk score is insufficient, because the cooling-off period a patient adversary must wait is a constant that does not grow with the horizon . We then present LoopHarness, which restores a persistent, non-decaying safety state at the loop level. Under mediated commits and an arbiter detection floor , it bounds the expected number of unauthorized irreversible actions by , a constant in , of which the term is decided by a model-free rule and therefore survives a fully colluding verifier. On native Agent-SafetyBench tasks, LoopHarness reduces overall outer-state attack success from 88.4–97.6% for the evaluated baselines to 0.1% while retaining 96.9% clean target completion; matched ablations, a controlled retention study, and an adaptive white-box red team further test the mechanisms underlying this result.