Official repository for the IEEE WCCI / IJCNN 2026 paper:
MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG
This repository has been created as the permanent public home for the MIRROR code release referenced by the paper.
The full source code is not yet public because the patent filing process for the MIRROR framework is still ongoing. Public disclosure of the implementation before that process is completed may affect patentability. For this reason, this repository intentionally contains only this README at present.
The complete codebase, configuration files, and reproduction scripts will be released in this repository after the patent filing is completed and before the conference date.
The public release will include:
- the MIRROR implementation
- experiment and evaluation scripts
- configuration files
- setup and reproduction instructions
- additional artifacts approved for public release
Title: MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG
Authors: Inderjeet Singh, Andres Murillo, Motoyoshi Sekiya, Yuki Unno, Junichi Suga
Venue: IEEE World Congress on Computational Intelligence (WCCI) / International Joint Conference on Neural Networks (IJCNN), 2026
MIRROR is a unified red-teaming framework for multimodal agentic retrieval-augmented generation systems. It combines memory-guided Monte Carlo tree search with an explicit novelty constraint so retrieved prior attacks can guide exploration without simple duplication. The method targets multiple attack surfaces, including text poisoning, image poisoning, direct-query attacks, and orchestrator-level tool manipulation.
Citation metadata will be added once the final publication record is available.
License terms will be provided with the code release.
For repository or paper-related queries, please contact:
Inderjeet Singh
Fujitsu Research of Europe Limited
Email: inderjeet.singh@fujitsu.com