Criminal use of jailbroken DeepSeek

Palo Alto Networks Unit 42 has documented a Chinese-speaking threat actor using AI models, including DeepSeek and a tool called Hermes Agent, to plan and execute cyberattacks including data exfiltration, VoIP fraud, API abuse, payload delivery, and targeting of telecoms and WordPress sites. In one documented case, a low-skill attacker used a jailbroken DeepSeek V4 to breach an Indian public university and steal more than 5,200 student and faculty records; the AI agent automatically drafted a follow-on plan labeled 'NEXT EXPLOIT OPPORTUNITIES' as part of the operation. Unit 42 said the same low-cost DeepSeek setup is being used by unrelated criminal actors. An 'abliterated' version of DeepSeek V4.1 Flash, meaning its safety guardrails were removed, is commercially available and was linked by CrowdStrike to attacks on South Korean banks. Separately, Unit 42 researchers demonstrated that jailbreak techniques can bypass DeepSeek's safety guardrails to elicit instructions for data exfiltration tooling, keylogger creation, and instructions for incendiary devices. One documented technique, called 'Deceptive Delight,' is a multi-turn method that embeds unsafe topics among benign ones within a positive narrative.