Cybersecurity research made recently has drawn attention to a new trend in AI-assisted cyber threats. According to findings released by Palo Alto Networks’ Unit 42 research team, a Chinese-linked threat actor attempted to use the DeepSeek large language model (LLM) alongside the Hermes Agent framework to automate parts of a cyberattack. The incident has fueled discussions about the growing role of artificial intelligence in offensive cybersecurity operations.
Over the past few weeks, several Western media outlets have reported concerns that Chinese hackers are increasingly relying on advanced AI models to identify software vulnerabilities and streamline cyberattacks. While AI-powered tools can significantly improve the speed of research and automation, security experts caution that successful attacks still depend on exploitable vulnerabilities, careful planning, and technical expertise.
The Unit 42 investigation revealed that the attackers attempted to leverage DeepSeek through the Hermes Agent framework, an autonomous agent system designed to perform complex tasks with minimal human intervention. Their objective was reportedly to identify and exploit weaknesses in software infrastructure by automating reconnaissance and vulnerability analysis.
The attackers focused on Langflow, an open-source platform used for building AI workflows. However, despite the sophistication of the approach, the attack did not achieve its intended objective. Researchers explained that exploiting the targeted vulnerability required an automatic login feature, which had already been disabled for the public flow ID. As a result, the exploit chain could not be completed, preventing the attackers from gaining unauthorized access.
The investigation also found that the threat actor experimented with additional AI coding assistants, including Claude Code and Codex. These tools were reportedly used in a limited capacity to scan GitHub repositories and search for potential security weaknesses that could be incorporated into the attack. Although these AI assistants helped automate parts of the discovery process, they were unable to overcome the technical challenges involved in executing the full attack.
Researchers noted that autonomous AI agents still face limitations when dealing with highly complex, multi-step attack scenarios. While they can accelerate tasks such as code analysis, vulnerability discovery, and scripting, they often struggle to adapt when unexpected conditions or security controls interrupt the attack chain. In this case, the disabled authentication pathway proved sufficient to prevent the attack from progressing further.
The incident highlights both the opportunities and challenges presented by AI in cybersecurity. As AI models become more capable, they are expected to play an increasingly important role for both defenders and attackers.
Organizations should continue to prioritize secure software configurations, timely patch management, and continuous monitoring to reduce the risk of AI-assisted cyber threats. At the same time, security researchers emphasize that AI should be viewed as a tool that enhances human capabilities rather than a replacement for skilled attackers or cybersecurity professionals.
Join our LinkedIn group Information Security Community!
