Tool dumps weights from TensorRT engines

Released by AI DRM startup RefortifAI, this tool proves that compiled engines are not a safe vault for intellectual property if the REFIT flag is enabled.

• –Uses a "refit-diff mapping" technique to identify exact file offsets for weights without needing the original model architecture or ONNX files.
• –Exports recovered weights to standard .safetensors or .npy formats, allowing models to be ported back to frameworks like PyTorch.
• –Highlights a major vulnerability in edge AI deployments where compiled .engine files are often incorrectly assumed to be obfuscated.
• –Restricted to engines built with BuilderFlag.REFIT, though this flag is commonly used for post-training quantization and runtime weight updates.
• –Serves as a "proof of concept" for why hardened runtimes are necessary for protecting proprietary weights in untrusted environments.