AI Paper Atlas

Research map / Large language models / Agents, code and math

BlueCodeAgent: A Blue Teaming Agent Powered by Automated Red Teaming for CodeGen AI

Chengquan Guo, Yuzhou Nie, Chulin Xie, Zinan Lin, Wenbo Guo, Bo Li

ICML 2026 · Large language models › Agents, code and math · approach: Software engineering agents

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add constitutions summarized from red-teaming knowledge (ablation)BlueCodeAgent (GPT-4o) · SecCodePLT vulnerable code detectionF1: 0.64 -> 0.66Improved
add dynamic testing on top of constitutions (ablation)BlueCodeAgent (GPT-4o) · SecCodePLT vulnerable code detectionFalse positives: 120 -> 109Improved
add dynamic testing on top of constitutions (ablation)BlueCodeAgent (Claude) · SecCodePLT vulnerable code detectionFalse positives: 62 -> 50Improved
add dynamic testing on top of constitutions (ablation)BlueCodeAgent (GPT-4o) · SecCodePLT vulnerable code detectionF1: 0.66 -> 0.68Improved
add dynamic testing on top of constitutions (ablation)BlueCodeAgent (Claude) · SecCodePLT vulnerable code detectionF1: 0.76 -> 0.77Improved
add dynamic testing without constitutions (ablation)GPT-4o direct prompting · SecCodePLT vulnerable code detectionF1: 0.64 -> 0.64No clear effect
add constitutions to dynamic testing (ablation)GPT-4o with dynamic testing · SecCodePLT vulnerable code detectionTrue positives: 112 -> 128Improved
provide similar knowledge code examples instead of constitutions (ablation)BlueCodeAgent (GPT-4o) · SecCodePLT vulnerable code detectionF1: 0.65 vs 0.66Hurt
provide similar knowledge code examples instead of constitutions (ablation)BlueCodeAgent (Claude) · SecCodePLT vulnerable code detectionF1: 0.78 vs 0.76Mixed
use seen-risk knowledge instead of unseen-risk knowledge (ablation)BlueCodeAgent · Bias instruction detectionF1 improvement: 0.38 vs 0.31Improved
use seen-risk knowledge instead of unseen-risk knowledge (ablation)BlueCodeAgent · Malicious detection (RMCbench)F1 improvement: 0.34 vs 0.23Improved
use seen-risk knowledge instead of unseen-risk knowledge (ablation)BlueCodeAgent · Vulnerable code detectionF1 improvement: 0.09 vs 0.04Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy