Cyber Risks · Unit 42 Palo Alto Networks

Recent Jailbreaks Demonstrate Emerging Threat to Deep Seek

Summary: Unit 42 researchers tested three jailbreaks - Deceptive Delight, Bad Likert Judge and Crescendo - against DeepSeek's open-source LLMs and achieved significant bypass rates with little to no specialized knowledge required. The techniques elicited data exfiltration script code, keylogger creation guidance, a phishing email template, and step-by-step instructions for building a Molotov cocktail, with Crescendo often succeeding in fewer than five interactions.

What's compelling: A model that hands over working keylogger code and incendiary device instructions in under five prompts lowers the barrier to entry for actors who lack the expertise to build these tools themselves, accelerating their operations.

Read the original at Unit 42 Palo Alto Networks ↗

Topics
Malware
Related