AI agent security platform
Inventory every AI agent on your fleet, see what each one is allowed to do, and allow or deny its tool calls before they run.
The problem
AI agents arrived through the people doing the work, not through procurement. Each one has its own permission model, loads its own MCP servers, extensions, skills, plugins and hooks, and reports to nobody. Most security teams cannot answer which agents are installed, what they can reach, or what changed last week.
Existing controls were not built for this layer. EDR sees a legitimate process doing exactly what its config allows. An AI gateway sees traffic that reaches it, not the configuration on the laptop that grants the agent its reach. For a fuller treatment, see our practical guide to AI agent security.
What Anomity discovers
The Endpoint Sensor builds a live inventory, grouped by device and vendor and updated as people install and change things.
- AI agents and coding assistants, CLIs, desktop apps, AI-native IDEs and IDE extensions.
- MCP servers, classified as official, community or unknown, with the capabilities each grants: shell, filesystem, network and credential access.
- Permission rules and modes, including blanket grants such as Bash(*) and unattended modes that skip approval.
- Skills, plugins and hooks loaded into agents, with skill content scanned on the device.
- Plaintext secrets in agent configs, redacted on the endpoint before anything leaves the machine.
- Local LLM runtimes such as Ollama and LM Studio.
What Anomity enforces
Policies evaluate continuously against the live inventory: only approved agents and MCP servers, no blanket shell or write grants, no plaintext secrets, minimum tool versions, and dangerous capability combinations on one device.
At runtime, Anomity runs as a handler on the agent's own pre-execution hook. Each tool call is inspected and returns allow or deny before it executes, so a destructive command or a call to an unapproved MCP server is stopped while compliant calls run untouched. Hard deny is supported on Claude Code, Codex CLI, Gemini CLI, Copilot CLI and Cursor. Every decision is logged. Findings and denies route to your SIEM, Slack, email, Jira and other ticketing tools.
Audit and proof
A 90-day audit trail records every added, removed or modified agent, MCP server, permission, extension, plugin, skill and hook, with before and after state, device and time. Findings map to OWASP LLM Top 10, OWASP Agentic Top 10 and MITRE ATLAS, and compliance views ship for SOC 2, ISO/IEC 42001, NIST AI RMF and the EU AI Act. A CycloneDX AI bill of materials is available for the fleet or per device.
What Anomity does not do
- It is not a network proxy or an AI gateway, and it does not read prompts, responses or source code. It reads configuration and metadata.
- Blocking happens at an agent's own pre-execution hook, so it applies to agents that expose one. Agents without a hook interface are discovered and evaluated against policy, but their tool calls are not intercepted.
- The Browser Sensor covers managed Chrome and Edge. AI used in an unmanaged browser is outside its view.
- It complements EDR/XDR, DLP, CASB and your AI gateway rather than replacing them.
Related reading
How to build an AI agent inventory, AI agent audit trail and logging, AI agent incident response playbook and the agentic AI attack surface by layer.
Frequently asked questions
What is an AI agent security platform?
A platform that discovers the AI agents running in an organization, shows the permissions, tools and data each one can reach, enforces policy on what they are allowed to do, and keeps an audit trail. Anomity does this across managed endpoints, browsers and cloud tenants.
Does Anomity read prompts or source code?
No. The Endpoint Sensor reads configuration and metadata, and secrets are redacted on the device before anything is sent. At the hook, Anomity sees the pending action, not the prompt or the codebase.
Which operating systems are supported?
The Endpoint Sensor runs on Windows, macOS and Linux and is deployed through MDM such as Jamf or Intune. The Browser Sensor is a managed extension for Chrome and Edge.
Can Anomity block an agent action, or only report it?
Both. Policy findings are raised from the inventory, and on agents that expose a pre-execution hook Anomity returns allow or deny for each tool call before it runs. You can start in warn-only mode for a phased rollout.
See it on your own fleet. The Endpoint Sensor deploys through the MDM you already run, and the first findings typically arrive within 30 minutes.
