SafeHarness: Lifecycle-Integrated Security Architecture for LLM Agent Deployment Harnesses
arXiv·high signal
New paper identifies the execution harness — the system layer orchestrating tool use, context management, and state persistence in LLM agents — as a high-value attack surface that existing security frameworks miss. SafeHarness proposes a lifecycle-integrated security architecture covering tool registration, context flow, and state persistence across the full agent deployment lifecycle. Directly relevant to anyone building or operating agent systems with tool access — the harness itself is the threat, not just the model.