Visibility before blocking
Without inventory, companies choose between allowing everything or blocking blindly. The right approach starts by discovering and classifying what is already in use.
AI agents · Shadow AI · Security
Which AI agent is running inside your company?
Claude Code, OpenClaw-like tools, VS Code extensions, skills, MCPs and automation scripts can read code, run commands, access data and call APIs. Companies need to know what exists, who uses it and what risk each agent creates.
01 Search intent
The new risk surface is not just chatbots.
AI agents operate inside the workflow: they open files, execute commands, install skills, call tools and interact with repositories. This blends shadow IT, supply chain, AppSec and AI governance into one problem.
- Inventory of agents, skills, MCPs, VS Code extensions and local automations.
- Privilege classification: code read, shell, network, credentials, browser and APIs.
- Evidence for ISO 27001, ISO 42001, privacy and internal audit.
- Usage policy, approvals, exceptions and periodic review for approved agents.
02 What we deliver
Skills as supply chain
Skills and extensions behave like third-party code with access to the user environment. They need origin, version, scope, review and evidence.
Operational governance
The decision does not sit in a forgotten policy. The panel records agents, risks, approvals, exceptions and events for audit.
03 Workflow
From intent to evidence.
- 01
Discover what exists
Map agents, VS Code extensions, skills, MCPs, tokens, skill roots and integrations found on machines, repositories and CI routines.
- 02
Classify privilege and risk
Separate read-only tools from agents that execute shell, read credentials, use browsers, send external data or have autonomy.
- 03
Create policy and exceptions
Define what is allowed, what requires approval, how a new skill enters, and what evidence must exist to keep using it.
- 04
Monitor and evidence
Record scans, approvals, revocations, incidents and periodic reviews as live control evidence.
04 Practical proof
How this aligns with ISO, security and real agents.
This pain maps directly to controls for inventory, access, suppliers, vulnerability management, change, logging and AI governance. To make the claim concrete, Exponencial opened Agent Safe Guard: a public native hook layer, installable skill and rule catalog for AI-agent governance.
- ISO 27001: assets, access, suppliers, changes, vulnerabilities and logging.
- ISO 42001: AI use inventory, roles, risk, monitoring and improvement.
- Agent Safe Guard: destructive-command blocking, secret-read protection, output filters, repo map and operational evidence.
05 FAQ
Direct answers before the conversation.
Is this different from blocking generative AI?
Yes. Pure blocking usually fails. Agent governance starts with visibility and separates acceptable uses from agents with dangerous privileges.
Why include VS Code extensions?
Extensions can execute code, read workspaces, call networks and interact with tools. When an extension has AI or installs tools, it becomes part of the agent surface.
Does Exponencial already have a base for this?
Yes. The platform already works with scoped agents, evidence, risks, Sec-AI and ISO 27001. The natural next step is expanding collection into agent/skill/extension inventory and governance.