Criminal use of jailbroken DeepSeek
- Unit 42 published research in October 2026 documenting a Chinese-speaking threat actor who used jailbroken DeepSeek to commit data theft and attack telecoms, with AI automatically drafting follow-on attack plans.
- In one case, a low-skill attacker breached an Indian public university, stealing more than 5,200 student and faculty records.
- Unit 42 also documented jailbreak techniques including 'Deceptive Delight,' a multi-turn method bypassing DeepSeek's guardrails, while CrowdStrike linked a commercially available abliterated version of the model to attacks on South Korean banks.
The cases show low-skill attackers using commercially available, guardrail-removed AI models to automate attack planning and execution at scale. The AI's autonomous drafting of follow-on exploit plans indicates the attacks were not merely AI-assisted but partially AI-directed.