BullyRAG:A Multi-Perspective RAG Robustness Evaluation Framework
summary
BullyRAG is an open-source RAG robustness evaluator that runs attacks out of the box against OpenAI, Anthropic and Hugging Face inference, using a weekly arXiv QA set (~1,000 triplets) and Berkeley function‑calling tests. It implements zero-width space, LTR/RTL and backspace retrieval poisons that drop cosine similarity from ~70 to ~20, plus five generation tricks (preferred keywords, self‑style, emotional stimuli, majority consensus, profit temptation) that induce malicious code execution or malicious links. A three-step Python setup reproduces successful cases (e.g., indices 0, 3, 4) and logs injected request.get + URL while still returning expected answers.
AI-written summary · about
chapters (10)
- 0:00BullyRAG framework intro
- 5:51What is RAG and its importance
- 9:25Robustness testing for RAG
- 12:00Attack objectives in BullyRAG
- 18:57Retrieval phase attacks
- 23:30Generation phase attacks
- 28:49BullyRAG datasets
- 33:48Using BullyRAG framework
- 40:53Key takeaways and open-source call
- 42:42Q&A
related talks
26mBreaking RAG Systems: Exploiting Vulnerabilities & Hardening Your GenAI Applicatio...
14mBeyond RAG: Building Reliable AI Systems for Privacy Assessments
05mopen-rag-eval: RAG Evaluation without "golden" answers
22mThe Power of RAG at Scale: A Year of Explosive Growth and Observab...
42mTake Your RAG To The Max!
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
30mEvaluating RAGs: On the correctness and coherence of Open Source eval metrics
22mCracking the Code: Deciphering Evaluation Essentials for RAG