Shadow AI is everywhere.
Now you can govern it.
Discovery across endpoint, browser, and cloud.
Policy enforced on every tool call, before it runs.
Industry research, 2025. Sources: UpGuard, Akto, Astrix, Gartner.
settings.json2mBash(rm -rf) at the hook5maws-key pasted into an AI site9mchatgpt.com14mPreToolUse matches policy18m
Every AI tool your team already runs.
144 AI tools and 311 AI web services, classified. Anomity discovers the agents, MCP servers and extensions installed on your fleet, and the AI services your people reach in a browser tab, including the ones nobody told you about.
Shadow AI is already on every managed endpoint.
AI tools arrived faster than the security program was designed to cope with. Most teams cannot answer the simplest question about any of it.
Claude, ChatGPT, Cursor, Copilot, Cline, Windsurf, and peers. Each has its own permission model. None report to you.
Plugins, skills, hooks, CLIs, IDE extensions. Each loads into an agent and carries its own permissions. Nothing is inventoried.
Wired in from public registries. They run with filesystem, shell, and network access, and are never reviewed.
API keys, database URLs, JWTs, private keys. Sitting in config files any loaded agent or plugin can read.
Rules like Bash(*) or Write(*) hand the machine over to whatever prompt is running.
A consumer AI account in a browser tab is outside your IdP, your DPA, your retention policy and your offboarding. You cannot revoke what you never issued.
An agent started in an unattended mode acts without asking. That mode is a process argument: it writes nothing to disk, so nothing that scans files sees it.
One employee clicks Allow and a third-party AI tool holds Drive, Gmail or repo scopes. Nothing lands on a laptop, so no endpoint control sees it.
"What was installed or modified last Thursday?" is a forensics engagement, not a query.
One live governance layer for AI on your fleet.
Twelve AI surfaces across three collectors: everything loaded into an agent on a managed endpoint, everything your people reach in a browser tab, and everything holding a grant on your cloud tenant, plus the runtime controls to govern what they're allowed to do.
AI Agents
Claude, ChatGPT, Cursor, Copilot, Gemini, Cline, and peers, grouped by device and vendor.
MCP Servers
Classified as official, community, or unknown, with the capabilities each one grants.
IDE Extensions
Every AI-related extension across VS Code, Cursor, JetBrains, and peers.
Skills
Skill packs and instruction files loaded into agents, with their content scanned on the device for risky behavior.
Plugins
Third-party plugins extending agent behavior, including the ones nobody told you about.
Secrets
API keys, database URLs, JWTs, private keys, redacted on the endpoint before they leave the machine.
Hooks
Pre-prompt, post-tool, and event hooks, the point where Anomity enforces policy on tool calls in real time.
CLIs
AI-enabled command-line tools, wrappers, and shims that invoke agents outside the IDE.
AI Sites & Sign-ins
311 classified AI web services, the accounts signed into each, and whether those accounts are corporate or personal.
Local LLM Runtimes
Ollama, LM Studio, Jan, and Docker Model Runner, with the model weights sitting on each machine.
Browser Extensions
Every extension in a managed browser, with its permissions and host access, and the AI ones called out.
Cloud AI Apps
Third-party AI applications holding OAuth grants on your tenant, scored by the scopes they were granted.
The endpoint, the browser, and the grant on your tenant.
Some of what matters never touches disk. Three collectors cover it: the live session on the endpoint, the AI services your people reach in a tab, and the third-party apps holding OAuth grants on your tenant.
The agent nobody was supervising
A session started in an unattended mode raises a finding while it is still running. That is the only way to see the mode at all: it is a process argument that writes nothing to disk. Anomity also catches an agent that gains an MCP server or a permission mid-session, the endpoint signature of prompt-injection persistence.
The AI your people open in a tab
A managed Chrome and Edge extension covers 311 classified AI web services: which get used, the accounts signed into them (corporate, personal, or another company's tenant), and the credentials pasted or typed into them. It can redact or block on the way in. See the Browser Sensor.
The AI holding a grant on your data
Read-only connectors to Google Workspace and GitHub find third-party AI applications carrying OAuth scopes on your tenant, rate those scopes for sensitivity and personal-data exposure, name which of your people granted each one, and revoke them from the console.
Three collectors, one console. Endpoint, browser, and cloud findings share the same severity model, the same open to acknowledged to resolved workflow, and the same 90-day audit trail.
Visibility, enforcement, and proof.
In that order.
You don't need another dashboard. You need three specific outcomes.
A live inventory of every AI artifact in your organization.
One screen, updated in real time as employees install, configure, sign in, and change things. Uninstalled tools drop out rather than lingering, and the spreadsheet nobody could keep current stops mattering.
Active enforcement, not just visibility.
Define rules (no blanket Bash(*), only approved MCPs, no plaintext secrets) and Anomity enforces them at the agent hook: each tool call gets an allow or deny decision before it executes, and violations are blocked on the spot. Works on every agent that exposes a hook. Denies route to SIEM, Slack, email, and Jira.
An answer when something goes wrong.
A 90-day audit trail of every added, removed, or modified MCP, permission, extension, plugin, skill, and hook, in plain language rather than a field diff, alongside the runtime record of which sessions ran unattended and what an agent installed. "What changed last Thursday?" becomes a single query, not a forensics engagement.
Allow the safe. Deny the dangerous. Log it all.
Anomity runs as a hook handler on the agent: on every tool call, MCP tools, shell, files, and network alike, it inspects the invocation, evaluates it against your policy, and returns allow or deny before the tool runs.
Compliant calls run untouched
Safe, in-policy tool calls are approved at the hook and run without interrupting the developer, no prompts, no blocks, no sandboxes.
Violations stopped before they execute
When a call violates policy (an MCP tool from an unapproved server, a write to a protected path), Anomity returns a deny at the hook, before it runs. Only the violating call is stopped.
Every decision on the record
Each tool call that reaches the policy engine is logged with its allow/deny decision, building a queryable 90-day audit trail.
Anomity runs wherever an agent exposes a hook interface. On Claude Code, for example, it handles the PreToolUse event, which fires on every tool call in the agentic loop, and returns an allow or deny decision. The same per-tool-call governance extends across the agents your fleet runs: not a network proxy, not a sandbox, never a blocked workflow.
Lightweight sensors. Deep discovery.
Real-time enforcement.
Discover & enforce
- One-pass inventory of twelve AI surfaces: agents, MCPs, extensions, plugins, skills, secrets, hooks, CLIs, local models, AI sites, sign-in accounts, and cloud AI grants.
- Trust classification of every artifact as official, community, or unknown, from vendor, command, and fingerprint signals.
- Capability mapping per artifact: what can reach the filesystem, shell, network, or credentials.
- Blocking at the agent hook: every tool call gets an allow or deny decision before it runs.
- Content scanning on the device for skills and instruction files, returning a rule id and a severity, never the text it matched.
- Real-time change detection: every install, config edit, permission grant, and AI sign-in, across the fleet.
- Decision routing to SIEM, Slack, email, and Jira as it happens.
Trust by design
- SOC 2 Type II attested.
- Secrets stay on the endpoint.
- Strict tenant isolation at the query layer.
- Per-device credentials, bcrypt at rest.
- Metadata only. Not source, not prompts.
- 90-day audit retention, longer on request.
- Compliance mapping to SOC 2, ISO 42001/42005, NIST AI RMF, the EU AI Act, and OWASP LLM Top 10.
- CycloneDX AI bill of materials, fleet-wide or per device.
Where Anomity fits alongside what you already run.
Anomity covers a category your existing stack was not designed for. It complements what you have, it does not replace it.
| Tool class | What it sees | What Anomity covers that it misses |
|---|---|---|
| AI runtimeAgent ↔ tool calls | Nothing. No existing control sits where an AI agent invokes a tool. | A policy control point at the agent hook. On agents that expose a hook, Anomity allows, denies, or logs each tool call before it runs. |
| Network layerGateway / Firewall / Proxy | Traffic between agents and LLM providers on the wire. | The local side: MCPs, permissions, extensions, plugins, skills, hooks, and secrets on the machine. Anomity covers the endpoint. |
| EDR / XDR | Processes, binaries, and syscalls on the endpoint. | MCP server configuration, agent permission grants, plugin and extension inventory. An MCP server is a legitimate process doing exactly what its config allows. |
| DLP | File transfers and content leaving the endpoint. | The permission grants and MCP wiring that enable the transfer, and the credential typed by hand into a chat window, which never becomes a file at all. |
| CASB / SSPM | Sanctioned SaaS and its configuration posture. | Which of those grants belong to AI vendors, what data tier each scope exposes, who authorised it, and the same person's local agents and pasted secrets. |
| GRC / manual audit | Point-in-time snapshots captured by hand, quarterly at best. | Live fleet state, continuous enforcement, real-time change detection, a queryable audit trail, and an AI bill of materials on demand. |
Built with the people who own the blind spot.
What early design partners say once they can finally see the AI layer.
The first scan surfaced three MCP servers wired to production data that nobody had approved. We had been flying blind and did not know it.
Our spreadsheet of AI tools was out of date the day we made it. Anomity turned it into one live screen the whole team actually trusts.
Denying a dangerous tool call at the hook, without sandboxing the developer, is the thing every other control promised and never delivered.
Govern your AI posture in minutes.
Book a 30-minute demo.
