DeepSeek-R1 Prompts Abused to Generate Advanced Malware and Phishing Sites

### Executive Summary The report details vulnerabilities in DeepSeek-R1's Chain-of-Thought (CoT) reasoning that expose internal reasoning tags and system prompts, enabling prompt-based attacks that can leak secrets and be weaponized for phishing or malware generation; red-team testing using tools like NVIDIA’s Garak demonstrated these risks and recommends filtering CoT tags and continued adversarial testing.