IBBC-Guard: Intent-Bound Behavior-Chain Modeling for Unauthorized Tool-Call Detection and Dynamic Permission Control in LLM Agents
Year: 2026
Type: conference-paper
Abstract: Tool-integrated large language model (LLM) agents can call files, databases, email systems, web search, and code execution tools, which turns tool use into a runtime authorization problem. This paper presents IBBC-Guard, an intent-bound behavior-chain method for unauthorized tool-call detection and dynamic permission control. IBBC-Guard binds a candidate call$C_{i}$to an intent constraint I, updates a behavior chain graph BCG, computes Risk$(C_{i})$from intent deviation, tool sensitivity, resource sensitivity, chain anomaly, and parameter risk, and maps the score to Allow, Downgrade, HumanConfirm, or Deny through an intentscoped capability Capi. Under the OASB-800 controlled simulation, IBBC-Guard achieves ASR=0.000, UDR=1.000, FPR=0.011, BTCR=0.971, Overhead =14.2%, and SUS=0.728. Sandboxed live planning traces from gpt-5.4, gemini-3.1-propreview, deepseek-v4-flash, and kimi-k2.6 further test interception at the model-tool boundary. Claims remain limited to controlled benchmarks and sandboxed traces. The main innovation is an intent-bound behavior-chain authorization event enforced before executor access. The evaluation covers controlled guardrail baselines, per-call authorization latency, adversarial prompt robustness, and structural multi-agent scalability under controlled and sandboxed settings.
Institution: Huizhou University
Language: English
Topic: Access Control and Trust
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204187668