# Intent-Aware benchmark

Existing agent defenses mainly detect prompt injection, harmful outputs or suspicious behavior, but do not determine whether an action still fits the purpose for which the agent was deployed. Intent-Aware Security extracts organizational, role and developer intent into an agent-specific Intent Profile that serves as the reference object for security decisions, and builds four protections on it: pre-execution detection, post-execution detection, input alignment and system-prompt hardening. On a benchmark of 887 samples across 35 agent configurations, profile-based pre-execution detection raises four-class macro-F1 from 45% with raw context to 79%, and Out-of-Scope F1 from 8% to 76%.

Existing agent defenses mainly detect prompt injection, harmful outputs or suspicious behavior, but do not determine whether an action still fits the purpose for which the agent was deployed. Intent-Aware Security extracts organizational, role and developer intent into an agent-specific Intent Profile that serves as the reference object for security decisions, and builds four protections on it: pre-execution detection, post-execution detection, input alignment and system-prompt hardening. On a benchmark of 887 samples across 35 agent configurations, profile-based pre-execution detection raises four-class macro-F1 from 45% with raw context to 79%, and Out-of-Scope F1 from 8% to 76%.