CodeTracer — forensic tool to trace backdoored AI-coding-assistant completions to poisoned training data

Researchers from University of Louisville and University of North Texas published CodeTracer (arXiv, week of 2026-07-16/20), a three-stage post-incident forensic pipeline that traces a harmful AI-coding-assistant code completion back to the specific poisoned training examples that caused it, achieving <0.03 false-negative rate across tested backdoor patterns.

First serious post-incident attribution capability for poisoned code-completion models, filling a gap between training-time data screening and output-time anomaly detection as AI-coding-agent supply-chain poisoning attacks increase.

AI/ML security research teams and coding-assistant vendors investigating suspected model poisoning incidents; not yet a packaged commercial tool.