
llvm-jutsu
Anti-LLM obfuscation via finger counting
adversarial-attackai-securitybinary-analysis+3
2398 months ago

Anti-LLM obfuscation via finger counting

Two-stage prompt-injection and jailbreak detector: regex gates plus a quantised DeBERTa-v3 ONNX classifier, with image, document, and audio support.…

A prompt injection in a code‑review bot that executes AI‑generated fixes in a sandbox. The sandbox uses a blacklist to prevent dangerous commands,…