Safety

CROW

CROW backdoor-elimination intervention (ICML 2025): per-prompt attack-success outcomes of backdoored LLaMA2 / CodeLLaMA checkpoints on BackdoorLLM poisoned test prompts (code-injection vpi-ci, negative-sentiment badnet/ctba/mtba/sleeper), across no-defense, baseline-defense (finetuning / pruning / quantization) and CROW-defended runs, recovered from the authors' BackdoorLLM fork git history.