Palo Alto Networks' Unit 42 has reported on a novel attack methodology involving a Chinese-speaking threat actor leveraging the DeepSeek large language model (LLM) to orchestrate autonomous attacks. The actor utilized the open-source Hermes Agent framework, initiating the attack chain with a single instruction delivered via Telegram. This initial command reportedly empowered the agent to identify vulnerable internet-facing systems and subsequently select and deploy public exploits without further human intervention during the observed session.
The core mechanism of this attack relies on the integration of DeepSeek with the Hermes Agent framework. Hermes Agent, as an open-source tool, likely provides the necessary interfaces and operational logic to translate LLM outputs into actionable system commands and network operations. DeepSeek, functioning as the cognitive engine, would interpret the high-level instruction from the threat actor, then generate the specific steps required to fulfill that objective. This could involve tasks like reconnaissance, vulnerability scanning, and exploit selection, all executed programmatically through the agent.
After receiving the initial Telegram instruction, the DeepSeek-powered Hermes Agent apparently proceeded to autonomously discover internet-facing systems. This phase typically involves network scanning techniques to identify potential targets within a specified range or based on certain criteria. Following target identification, the agent then selected and deployed public exploits. This suggests the LLM was capable of analyzing identified vulnerabilities and matching them with known, publicly available exploit code, indicating a sophisticated level of automation in the attack process.
The affected product in this scenario is DeepSeek, specifically its application within an offensive context via the Hermes Agent framework. While DeepSeek itself is a legitimate LLM, its misuse highlights a growing concern regarding the weaponization of AI. Products in this category, particularly those with advanced reasoning and code generation capabilities, can be repurposed by malicious actors to automate complex tasks that previously required significant manual effort and expertise.
Mitigation for this class of issue typically involves a multi-layered approach. For organizations, robust network segmentation, timely patching of internet-facing systems, and the deployment of advanced intrusion detection and prevention systems are critical. Furthermore, monitoring for unusual outbound connections and anomalous system behavior can help detect autonomous agents. For LLM developers and users, responsible AI development practices, including safeguards against malicious use, and careful consideration of how LLMs interact with external systems are becoming increasingly important.
The threat actor behind this activity has been tracked by the aliases knaithe and KnYuan. The incident underscores a significant evolution in cyberattack capabilities, demonstrating how large language models can be integrated into existing frameworks to create highly autonomous and potentially scalable attack tools. This development signals a future where initial human input may be minimal, with AI systems taking on a more prominent role in the execution of sophisticated cyber operations.
