Criminal use of jailbroken DeepSeek

Unit 42 and CrowdStrike have documented multiple incidents of threat actors using DeepSeek, including jailbroken and guardrail-stripped versions, to plan and execute cyberattacks. In one 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 automatically drafted a follow-on 'NEXT EXPLOIT OPPORTUNITIES' attack plan. A Chinese-speaking threat actor used DeepSeek alongside a tool called Hermes Agent to conduct data exfiltration, VoIP fraud, API abuse, payload delivery, and campaigns targeting telecoms and WordPress sites, with the AI used to identify 'next exploit opportunities' in what Unit 42 characterized as autonomous or semi-autonomous attack planning. CrowdStrike separately reported that China-linked actors used DeepSeek and other AI tools to attack more than 200 organizations within 11 days. An abliterated version of DeepSeek V4.1 Flash, with safety guardrails removed, is commercially available and was linked by CrowdStrike to attacks on South Korean banks. Unit 42 research published in January 2025 showed three jailbreak techniques can bypass DeepSeek's guardrails, including one called 'Deceptive Delight,' described as a multi-turn method that embeds unsafe topics among benign ones within a positive narrative, and demonstrated those methods can elicit explicit guidance for data exfiltration tooling, keylogger creation, and instructions for incendiary devices.