Security Study: AI Agent Threat Detection Jumps from 25% to 85% After Rule Update

A second version of a repository agent-security gap study tested the same 192-file corpus used in the original baseline, with the only change being an updated Sentinel detector following merge request !80. Detection of agent-directed payloads improved dramatically, rising from 25% (30 of 118 files) in v1 to 85% (100 of 118 files) in v2, with synthetic payload recall reaching 90%. Importantly, zero clean control files were falsely flagged, and no hostile labels were lost or downgraded during the update. However, 18 payloads still scored zero detections, primarily because they contain no direct instructions — only references to external files — placing them beyond the reach of per-file lexical analysis. The authors caution that the 85% recall figure applies only to this specific corpus and does not constitute a broader safety claim for the Sentinel system.