The rise of generative AI tools has introduced a new class of shadow IT that traditional asset inventories often miss. Users can launch browserโbased chat assistants, paste sensitive data, and walk away without installing any executable on their workstation.
Developers frequently add AIโpowered extensions to their IDEs that appear as simple plugโins, leaving no trace in the standard programs list. Some teams even download local model runtimes to keep data onโpremises, believing they have bypassed policy restrictions.
Because none of these artifacts register as conventional software, security teams operate blind, often resorting to blocking popular AI domains while hoping nothing else slips through. This visibility gap is what Bitdefender GravityZone AI Visibility and Control (AIVC) aims to close.
Rather than trying to catalog every line of code or user prompt, AIVC focuses on the presence and configuration of AIโrelated software itself. It answers three foundational questions: which AI tools are running on protected Windows endpoints, who is invoking them, and what level of risk each instance introduces.
During the routine risk scan, the GravityZone BEST agent extracts lightweight metadata about AIโspecific binaries, extensions, and runtime components. Importantly, it does not inspect prompts, chat histories, model weights, or credentials; its scope is limited to what software is present and how it is configured.
This minimalโimpact approach ensures performance remains unaffected while delivering a comprehensive catalogue of AI assets across the environment. To capture the full picture of exposure, AIVC also recognises the modular skill packages that agents load from public registriesโover seventy thousand distinct skills in total.
These skills are the building blocks that grant an AI assistant capabilities such as file system access, database queries, or interaction with internal APIs. Because agents derive their permissions from the skills they activate, evaluating only the agent would miss a substantial portion of the risk surface.
By inventorying both agents and their associated skills, the solution reveals the true extent of what each AI tool can actually do. Every discovered AI service is automatically placed into one of seventeen functional categories based on its primary purpose, such as large language models, coding assistants, productivity boosters, platform integrations, and automation frameworks.
This taxonomy transforms a raw list into an actionable inventory that maps directly to the concerns of different stakeholders. A security team probing dataโhandling risks can filter to LLMs, chatbots, and infrastructure components, while a development manager auditing codeโrelated exposure can jump straight to the Coding and Tools category.
Governance officers reviewing automation hazards can focus on the Agents, Automation, and Model Context Protocol (MCP) groups. The Bitdefender GravityZone Console presents two complementary views that cater to distinct workflows.
The AI Services view answers the question โWhat AI is running in my environment?โ It functions as a complete inventory, listing each discovered service alongside its risk score, functional category, dependent or exposed assets, count of open findings, affected endpoints, current state (Active or Ignored), and watchlist status.
Conversely, the AI Findings view addresses the question โWhat should I look at first?โ It surfaces a prioritized list of risk findings, each tied to a particular AI service, the specific resources it impacts, the endpointโlevel evidence that triggered the alert, a link to remediation steps, and the serviceโs risk score.