CanaryCap-Stego: Supplementary Data and Analysis Package

This dataset accompanies the manuscript “CanaryCap-Stego: Steganographic Capability Canaries for Detecting Agentic-AI Tool Misuse and Hidden Retrieval Abuse.” It contains the de-identified experimental data and analysis artifacts used to support the reported evaluation of CanaryCap-Stego across agentic-AI security scenarios. The package includes experimental trajectory matrices, carrier and scenario manifests, analysis configurations, statistical outputs, runtime measurements, transformation and capacity results, ablation and adaptive-stress results, generated figures, source-to-output traceability records, reproducibility documentation, and integrity checksums. These materials support analysis-level verification and regeneration of the statistical summaries, tables, and figures reported in the manuscript. The dataset covers the study’s confirmatory evaluation of 600 carriers, 2,700 scenarios, three model families, and 127,800 agent trajectories, together with supplementary transformation, capacity, runtime, and robustness analyses. The package does not constitute a complete replay environment for provider-hosted model calls and does not include the operational CanaryCap-Stego source code. The source code contains security-sensitive carrier-positioning, attack-generation, and misuse-response logic and is available from the corresponding author upon reasonable request for bona fide academic reproducibility or security review.