Risk-Adaptive Continuous Attestation for AI Agents

Abstract

AI agents can perform powerful actions — reading files, sending emails, transferring funds, or deleting databases. Most security systems either check every action the same way (wasteful) or skip checking entirely (dangerous). We propose Risk-Adaptive Continuous Attestation (RACA), a security middleware that decides when to perform fresh attestation checks based on the risk level of each tool action. LOW and MEDIUM risk actions reuse existing security tokens. HIGH and CRITICAL risk actions always require a fresh check — no exceptions. We evaluated RACA in two ways:- Experiment 1: Simulation with 15,000 trials using LangGraph — 100% attack detection with 50.2% fewer security checks- Experiment 2: Real LLM experiment with llama3.2 across 1,000 trials — 100% attack detection with 49.5% token reuse Results are consistent across both experimental setups, confirming that matching security checking frequency to action risk gives the same protection as always-checking with half the overhead. Included in this upload:- Full research paper (PDF)- Core RACA implementation and experiment code- Simulation results dataset (15,000 trials)- Real LLM results dataset (1,000 trials)- All figures from the paper GitHub Repository: https://github.com/Shahreyaarr/Risk-adaptive-attestation-ai-security