Answers are grounded in this paper and cite their sources.
As autonomous large language model (LLM) agents gain tool-use and long-horizon planning capabilities, they present a growing risk of executing harmful actions without explicit human authorization. Existing safety measures—such as alignment training, prompt hardening, and execution-time moderation—often fail against multi-step adversarial strategies in which an initially benign request is progressively transformed into a dangerous outcome. We introduce Kaifeng Prefecture, a safe-harness architecture that prevents autonomous LLM-agent attacks through structural separation of roles and control-flow inversion. The architecture assigns evaluative, observational, predictive, and capability functions to distinct modules (Value Judge, Observer, Predictor, and Capability), routes high-risk actions through an external gate, and preserves a human interrupt channel. Three formal propositions and accompanying floor invariants characterize the conditions under which harmful commands are blocked before execution. To make the design concrete, we describe a 23-scenario prototype implemented with a TranscriptBackend, demonstrating how the harness intercepts and neutralizes attack chains that bypass conventional guardrails. The result is a safety architecture that limits the autonomy of LLM agents without requiring full symbolic verification of every possible plan.
No takes yet. Share an insight, caveat, or question.
Fan He (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: