Prompt Steel: Russia-linked threat group uses LLM-powered tool to automate data theft

A Russia-linked group is reported to be using a tool called Prompt Steel that relies on a large language model to generate commands for stealing information, illustrating the operational use of AI agents in cyber attacks against organizations including SMBs.[2] The briefing highlights Prompt Steel as part of a broader surge in AI-driven cyber threats, with evolving malware tactics and growing risk of automated data theft enabled by LLM-generated instructions.[2]

The report describes a Russia-linked group using an LLM-powered malware tool (Prompt Steel / PROMPTSTEAL-type capability) that queries large language models via APIs (e.g., Hugging Face) to dynamically generate Windows commands for reconnaissance and data theft during live operations.[6][10] This reflects an operational use of AI agents within malware, where the model is effectively an on-demand decision and command-generation component rather than just a pre-attack productivity aid.[6][9][10] From a RealGround perspective, this exemplifies AI agent abuse: adversaries are wiring LLMs into autonomous attack loops that can adapt commands, evade static detection, and scale automated data theft against SMBs and larger organizations. Practically, defenders need to treat LLM backends and their APIs as part of the attack surface, applying secure AI agent design, continuous AI-focused red teaming, and business-logic audits to detect and constrain any agent-like components that can issue system, network, or data-access commands.