Megadose AI progress, ranked and analyzed.

CodePoisonRAG: Knowledge Poisoning Attacks on Retrieval-Augmented Code Generation

· ArXiv · AI/CL/LG ·
A single planted code artifact was enough to steer retrieval-augmented generators toward specific vulnerabilities.

The paper describes CodePoisonRAG, a black-box poisoning method that turns benign fixed-code entries into task-matched vulnerable artifacts with false safety claims. In tests, 85 poisoned artifacts across ten CWE classes in Java and C all landed in the Top-3 retrieval results for their target queries. Across three generators, the reported attack success rate was 0.80 to 0.93. A defense called CodeGuarder reduced but did not stop the attack, with success rates still at 0.40 to 0.71. ArXiv · AI/CL/LG's note

score 5

Categories: Research