Guides · ← All guides
How to Find Shadow AI Agents Running in Your Company
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer
Shadow agents surface through their footprints, not their code: API keys with no owner, spend lines nobody recognises, outbound traffic to model providers from unexpected hosts, and SaaS audit logs showing integrations nobody approved. Sweep those four sources, reconcile them against the agent register, and every unreconciled item is either a shadow agent or a credential leak — both worth finding. The fix is cultural as much as technical: a register, a one-week path to a legitimate scoped agent, and an amnesty for declaring existing ones. You cannot apply permission scoping to an agent you do not know exists, which is why shadow discovery precedes every other control in a maturing fleet.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works while you read — most controls in this guide start working the moment a canary, a policy check, or a spend cap is in place.
What is a shadow AI agent? any automated system that reasons and acts with company data or credentials outside the registered inventory — built by an employee, a contractor or a forgotten project, unknown to whoever owns agent risk.
The four footprint sources
Provider-side records first: your model and API providers can list keys, organisations and usage. Every key maps to a person and a purpose, or it is a finding. Finance second: card spend and invoices for AI tools by department — the $20-a-month seats are not the risk, but the automation platforms and API top-ups are a census written by accounting. Network third: egress logs showing which internal hosts talk to model APIs and automation webhooks; a server that reasons will eventually call home, and the call is in the proxy whether or not the project was.
SaaS audit logs fourth: the CRM, mail and document platforms record which integrations and tokens read and write. An unknown integration with export scope in the CRM is a shadow agent until proven otherwise — and its read pattern in those logs usually reveals what it is for. Cross-reference all four against the register from the compliance pack: the register is the claim, the footprints are the evidence.
Where they hide, concretely
The greatest hits repeat across companies. A workflow platform account under a personal email, automating a real process with the founder's all-access token — the n8n pattern from the n8n checklist, minus the checklist. A channel bot in a team chat, summarising threads into a spreadsheet nobody owns. A script on a workstation, scheduled, holding a production key, written by someone who left in March. A browser-side assistant reading the CRM through an employee's session, invisible to API logs and visible only in the CRM's own access records.
Contractors deserve a dedicated sweep: deliverables get handed over, but the automations built during the engagement often keep running on credentials issued for the project. Exit checklists should name agent handover explicitly — keys returned, workflows transferred or killed — or the engagement never really ends.
Amnesty and the legitimate path
Discovery framed as enforcement produces concealment; the second sweep finds fewer agents and the real number grows. Run the first sweep as an amnesty: declare by a date, no blame, and the team gets help migrating the agent onto the register with scoped credentials and a policy boundary. Most shadow agents exist because someone had real work and no sanctioned route to automate it — punishing the declaration teaches the next person to bury theirs deeper.
Pair the amnesty with a genuinely fast legitimate path: a registered agent with scoped keys inside a week. If the sanctioned route takes a quarter, the shadow route wins on merit and everyone in the building knows it. The register's growth after an amnesty is the metric — a census that finds three agents in a 200-person company has not found the agents; it has found the honest.
Disposition: register, rebuild or retire
Each found agent gets one of three dispositions, recorded. Register: the agent is sound and needed — scope its credentials, put it behind the policy boundary, plant its canaries, name its owner. Rebuild: the work is needed but the implementation is a liability — standing admin keys, no logs, a personal account — so it is re-created properly, in the open. Retire: the work stopped mattering or the risk never justified it — kill the keys first, then the workflow, then confirm the footprints stop.
The order inside retire matters: credentials before workflows, because a deleted workflow with a live key is a shadow agent waiting for a new host. Confirm retirement in the footprint sources a month later; a key that still calls after its funeral is the most important finding the second sweep can make.
Keeping the fleet known
One sweep is a photograph; make it a subscription. New API keys require a register entry before they activate. Provider spend alerts route to the register's owner. Quarterly, the four footprint sources reconcile automatically and only the exceptions consume human time. Rogue-Agent Defense's fleet view is built around this discipline for mediated agents — the point generalises: the inventory is a living control, not an annual document.
And keep the door open permanently: any employee can register an agent in minutes, with the security conversation happening as enablement — scoping, canaries, approvals — rather than as permission-seeking. The companies that know their fleets are not the ones with the strictest bans; they are the ones where declaring an agent is easier than hiding one.
Five actions to take today
- Pull the four footprint sources. Provider keys and usage, finance spend, proxy egress to model APIs, SaaS integration logs.
- Reconcile against the register. Every unmatched footprint is a shadow agent or a leaked key — open a line for each.
- Declare an amnesty window. Register by the date, no blame, migration help included — and a one-week legitimate path behind it.
- Disposition every find. Register, rebuild or retire, recorded; on retire, kill keys before workflows and verify silence a month later.
- Wire registration into key issuance. No new agent credential activates without a register entry and a named owner.
A worked example: the census at Halvern Manufacturing
The following is a fictional example built from real threat classes and controls. The company and numbers are invented; the mechanics are not.
Halvern Manufacturing, a 210-person firm, believes it runs two agents — a registered support assistant and a registered reporting workflow. The first footprint sweep takes its IT lead three days: provider records show 9 active API keys against 2 registered, finance finds four automation subscriptions nobody claimed, and the CRM audit log shows an unknown integration reading the full contact table nightly.
The amnesty does the rest. Declared within the window: a quoting workflow on a personal automation account holding a sales manager's token; a channel bot archiving engineering threads; a contractor's sync from the previous spring, still running — that one is the CRM integration, built for a data-cleanup project and never switched off. Undeclared but found by footprint: a workstation script summarising supplier emails with a production mail key, written by an employee who had left two months earlier.
Dispositions: the quoting workflow is rebuilt on the register with scoped keys; the channel bot is registered with its archive access narrowed; the contractor sync and the orphaned script are retired, keys first. Final count: 7 real agents, 5 registered or rebuilt, 2 retired — and a quarterly reconciliation that now takes the IT lead an hour, because the exceptions list starts empty and mostly stays that way.
Frequently asked questions
What is a shadow AI agent?
Any system that reasons and acts with company data or credentials outside the registered inventory: an employee's workflow on a personal account, a contractor's leftover automation, a scheduled script holding a production key. The risk is not that it is malicious — it is that no scoping, logging, canaries or owner apply to it.
How do you find shadow AI agents?
Through footprints, reconciled against the register: model-provider keys and usage, finance spend on AI tools, proxy egress to model APIs and automation webhooks, and SaaS audit logs showing integrations and their read patterns. Every unmatched footprint is a shadow agent or a leaked credential until reconciled.
Why do shadow agents appear?
Because someone had real work and no fast sanctioned route to automate it. Personal accounts, founder tokens and leftover project credentials fill the gap. The durable fix pairs discovery with an amnesty and a legitimate path measured in days — if the sanctioned route takes a quarter, the shadow route wins.
What should happen to a shadow agent once found?
One of three recorded dispositions: register it (scope credentials, add the policy boundary, name an owner), rebuild it (the work is needed, the implementation is a liability), or retire it (kill the credentials first, then the workflow, then verify the footprints stop a month later).
How do you keep the agent inventory complete over time?
Make registration a property of issuance: no new agent credential activates without a register entry and owner. Reconcile the four footprint sources quarterly — automatically, with humans seeing only exceptions — and keep declaration permanently easier than concealment.
Related guides
- AI Agent Compliance Evidence — the register the footprints reconcile against.
- Agent Permission Scoping — the control shadow agents bypass entirely.
- n8n AI Agent Security Checklist — the most common shadow habitat, hardened.
- Where to Plant Canary Tokens — tripwires for every newly registered agent.
- How to Rotate Credentials After an Agent Compromise — for the keys the census finds ownerless.
- AI Agent Velocity Anomalies — footprints as a continuous signal.
Generate a free canary token → No signup, nothing stored: plant a tripwire where your agent works, and let the next rogue action announce itself.