ECIR: Evidence-Calibrated LLM Security Assessment Agent
Year: 2026
Type: preprint
Abstract: ECIR (Evidence-Calibrated LLM Security Assessment Agent) is a research framework for evaluating and calibrating security assessment decisions made by large language model (LLM) agents under uncertain, incomplete, and potentially adversarial evidence. The framework models security assessment as a sequence of evidence interpretation, hypothesis generation, testing, verification, and decision-making stages, with particular emphasis on the ACT, WAIT, and ESCALATE decision boundary. ECIR introduces an evidence-trust-aware evaluation protocol designed to measure not only whether an agent reaches a correct security assessment, but also whether its escalation behavior is appropriately calibrated to the reliability and completeness of available evidence. This release contains the research artifacts associated with the ECIR study, including the framework implementation, evaluation scenarios, experimental configuration, and reproducibility materials. This work is intended for research and authorized security assessment only.
Author: Ibrahim Zain Al-Sabreen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204945853
Cites:
Cited by:
Related to: