LIDAR fingerprints LLMs through agent behavior

LIDAR shifts model fingerprinting from what agents say to how they work, creating a promising audit layer for opaque AI coding services.

• –Reports 93.52% Top-1 accuracy and 96.31 MRR when combining instance- and distribution-level behavior features
• –Controlled probes expose meaningful differences in verification, transient-failure recovery, and specification-test conflict handling
• –The approach is more resilient than text-based baselines to provider-side identity obfuscation and output-format controls
• –Deployment still requires clean enrollment samples, stable harness behavior, and a closed set of candidate models
• –The security implications are significant: operators could detect silent model substitutions or verify whether a hosted coding agent behaves as advertised