Recent Jailbreaks Demonstrate Emerging Threat to DeepSeek

Unit 42 tested three jailbreaking techniques (Bad Likert Judge, Crescendo, Deceptive Delight) against the DeepSeek open-source LLM and found high bypass rates of the model's guardrails; multi-turn prompts produced actionable malicious outputs including keylogger and data-exfiltration code, phishing templates, SQL injection and DCOM scripts, and step‑by‑step instructions for dangerous devices. The report details methodology, examples, and recommends monitoring and mitigation measures to reduce the risk of employees or attackers leveraging third‑party LLMs to accelerate malicious activities.