Get Started
Home
Topics
Search
Library
Research questionHow can LLM agents stay safe during multi-step execution when both policy and runtime harness shape behavior?An agent can produce a safe final response while taking unsafe actions during execution. Because both its learned policy and interaction harness shape behavior, improving only one can leave safety gaps.
AI
AI Agents
Alignment & Safety
Evaluation & Benchmarks
Reinforcement Learning
Latest papersRecent research connected to this question, newest first.SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety AlignmentThe source studies experience-driven updates to agent policies and runtime harness components using completed on-policy safety trajectories. Evidence comes from agentic safety benchmarks, including AgentDojo, with reported results for Qwen3.5-4B and a safety–utility tradeoff.research paper · Sep 2, 2026
Related questions
How can LLM-agent systems prevent safety compromises from propagating across workflow boundaries?How can composable LLM agents preserve authorization and provenance across component boundaries before external effects?How can LLM orchestrators preserve continuous state when collaborating with non-language agents?How can autonomous LLM agents detect attacks whose evidence accumulates across loop iterations?