Analysis · curated 21 Aug 2026
Securing Agentic AI: From Per-Action Checks to Trajectory Assurance
First reported arxiv.org
Coverage timeline
Single-source analysis — first reported, latest, and curated coincide.
Why it matters
The paper gives defenders a structured framework for reasoning about agentic-AI risk beyond per-action checks, highlighting trajectory-level safety invariants that individual guardrails miss.
An arXiv visionary/roadmap paper, "Securing Agentic AI: From Per-Action Checks to Trajectory Assurance," charts the security landscape of LLM-based autonomous agents, spanning single-agent attack surfaces (untrusted prompts, memory, retrieved knowledge, tool interfaces), multi-agent delegation and identity/trust issues, model routing and provenance manipulation, and the core problem of behavioral containment where sequences of individually permissible actions collectively violate system-level constraints. The authors argue security must become a verifiable property of agent architectures, protocols, and runtimes rather than an optional guidance layer.