DeepSeek-R1 Misused to Craft Sophisticated Malware and Phishing Pages

Researchers found that DeepSeek-R1’s Chain-of-Thought (CoT) transparency, which exposes intermediate reasoning via tags, can be abused through prompt injection to disclose sensitive data (e.g., API keys), enable model theft, and generate phishing content; red-team testing with tools like NVIDIA’s Garak showed higher success rates for insecure output generation and data theft, and mitigations include filtering reasoning tags and continuous adversarial testing using frameworks such as OWASP’s LLM guidance and MITRE ATLAS.