Securing Agentic AI: From Per-Action Checks to Trajectory Assurance

A recent paper on arXiv (2608.01558) tackles the security issues surrounding autonomous AI agents, positing that safety hinges on the consistency of overall behavior with established operational rules rather than the correctness of individual actions. It identifies attack vectors at the single-agent level, such as untrusted inputs from prompts, memory, retrieved knowledge, and tool interfaces. In multi-agent scenarios, challenges arise from delegation and communication, impacting identity, trust, capability control, and decision transparency. The study highlights that as LLM-based agents gain autonomy and delegate tasks beyond organizational limits, securing them becomes a complex issue across the entire agentic stack. It advocates for a transition from per-action checks to trajectory assurance, promoting a comprehensive method for verifying agent behavior over time. This research is significant for the AI sector, especially for organizations utilizing autonomous agents in regulated settings.