Frontier artificial intelligence labs log flagged user prompt interactions to train proprietary internal models on security vulnerabilities