What is a shadow AI agent?
An AI agent plans and acts on its own: it calls tools and APIs, sends emails, creates tickets or edits files. That autonomy is a design choice, not a governance failure. An agent becomes a shadow AI agent when it operates outside the organization's visibility, approval or governance: no inventory entry, no owner, no review of what it is allowed to do.
Shadow agents often run with the permissions of the person who set them up, and they can keep running after that person changes teams or leaves. That sets them apart from shadow IT and shadow AI in the usual sense, where an employee uses an unapproved app or chatbot and stays in control of every step. The categories overlap rather than exclude each other: a shadow AI agent is one form of shadow AI, and it often runs on a shadow IT platform.
| Shadow IT | Shadow AI (chat use) | Shadow AI agent | |
|---|---|---|---|
| What it is | Unapproved app or cloud service | Unapproved use of an AI chat tool | Unapproved AI agent that plans and carries out tasks |
| Acts on its own | No | No, a person prompts each step | Yes, within its tools and permissions |
| Holds credentials | The user's account | The user's account | Often a user's OAuth token or a shared service account |
| Runs without the user | No | No | Yes, on triggers or schedules |
| Typical risk | Data stored outside approved systems | Confidential data pasted into prompts | Actions taken with excess rights, by nobody's decision |
The last column is why chat-use policies are not enough. A policy that says "don't paste customer data into public chatbots" says nothing about an agent that already has read and send rights on a sales mailbox.
Where do shadow AI agents come from?
Most shadow agents are not built by attackers or rogue developers. Employees create them to save time, usually through one of these routes:
- agent features switched on inside SaaS tools the company already licenses;
- low-code and automation builders that let business teams chain AI steps with mailboxes, CRMs and file shares;
- browser assistants and extensions that read and act on web pages;
- coding agents with access to repositories and build pipelines;
- connectors (for example MCP servers) and personal OAuth grants to mail, calendar or drive;
- customer-service agents built in business-unit projects outside central IT.
Example scenario (fictional): A sales rep connects an AI agent to their mailbox and the CRM so it drafts follow-ups and updates deal stages. The rep leaves the company. IT disables their account in the identity provider. Whether that also stops the agent depends on how its access was granted: a long-lived API key, or a token issued by a service outside the company's single sign-on, can keep working until someone revokes it explicitly. In this scenario the agent keeps running, and nobody revokes its access, because nobody knows the agent exists.
What are the security risks of shadow AI agents?
The core risk is that an unreviewed agent takes real actions with real credentials while nobody watches. In practice, that shows up in five ways:
- Excess and inherited permissions. Agents often get the full rights of their creator's account or of a shared service account, far more than their task needs.
- Orphaned access. When the creator leaves, the agent's tokens or keys can survive offboarding unless someone revokes them.
- Prompt injection. An agent that reads emails, documents or web pages can follow instructions hidden in that content and act on them with its permissions.
- Unchecked approvals. People tend to approve confident-sounding agent actions without checking them.
- Compliance blind spots. Nobody can check whether an agent processes personal data or falls into an AI Act high-risk area if nobody knows it exists.
The three steps below address these risks in order: discovery and an inventory make agents visible, ownership closes the orphan gap, and least privilege limits what any agent can do, including a hijacked one.
Step 1: How to discover shadow AI agents and build an inventory
Shadow AI detection starts where agents leave traces. No single tool sees every agent, so combine several sources:
- OAuth app grants and consented third-party apps in your identity provider;
- admin consoles of the SaaS tools you license, which often show which AI or agent features are switched on and by whom;
- proxy, secure web gateway or CASB logs showing traffic to AI services and automation platforms;
- the admin views of low-code and automation platforms;
- a self-reporting call to business teams.
An API inventory will not show an agent that lives entirely inside a low-code platform, so treat self-reporting as a discovery source in its own right. Make it an amnesty: teams that register an agent in the first weeks keep it running while it is reviewed.
Once you find an agent, record it. An AI agent inventory is a register that records, for every agent, what it does, whose authority it acts on and what it can reach. These fields make the register useful for security reviews and for AI Act questions later:
| Field | Why you need it |
|---|---|
| Name and ID | Unique reference for tickets, logs and audits |
| Platform | Where the agent runs (SaaS feature, low-code tool, own code) |
| Business purpose | Lets you judge whether its permissions fit the job |
| Business owner | Person accountable for what the agent does |
| Technical owner and deputy | Who configures it, rotates credentials and can switch it off |
| Identity type | Own agent identity, a user's token or a shared service account |
| Permissions and scopes | What it can read, write, send or delete |
| Data classes | Personal data? Confidential data? Determines privacy review |
| Connected tools and connectors | Every system it can reach, including MCP servers and plugins |
| Autonomy level and approval points | Which actions need a human click first |
| External exposure | Does it talk to customers or other people outside the company? |
| Last review and expiry date | Forces a decision instead of indefinite running |
Step 2: Give every agent an owner
Every agent needs a named business owner and a technical owner, plus a deputy for each. The business owner answers for what the agent does; the technical owner answers for how it does it. The other roles are consulted or informed:
| Activity | Business owner | Technical owner | Security / IAM | Data protection | AI / compliance lead | Works council |
|---|---|---|---|---|---|---|
| Define purpose and accept results | A, R | C | I | C | I | |
| Configure, manage credentials, rotate tokens | I | A, R | C | |||
| Approve permissions and monitoring | C | R | A | I | ||
| Review personal-data processing, check whether a DPIA is needed | C | C | I | A, R | I | |
| Classify under the AI Act | C | C | I | C | A, R | |
| Agents that process employee behavior or performance data | C | C | C | C | C | Co-determination where applicable |
A = accountable, R = responsible, C = consulted, I = informed.
Two practices make the model hold. First, an agent without an owner gets suspended or moved into a restricted environment until someone claims it. That is a house rule, not a legal requirement, but it is the only way to stop the inventory from filling up with orphans. Second, add agents to offboarding: when an owner leaves, their agents and tokens are handed over or switched off, just like their laptop.
For the works council column: in Germany, where a works council exists, it has a co-determination right over technical systems designed to monitor employee behavior or performance (§ 87(1) no. 6 BetrVG). German labor courts read "designed to" broadly: a system that is objectively capable of monitoring can be enough, whatever the employer intends. Whether a given agent falls under that depends on what data it processes and how; counsel should assess. This is not legal advice.
Step 3: How to secure AI agents using least privilege
The principle is least agency, which extends least privilege from permissions to tools and autonomy: give an agent only the tools, permissions and autonomy its task needs. The OWASP Top 10 for LLM Applications 2025 names three root causes of "excessive agency": excessive functionality, excessive permissions and excessive autonomy. Its countermeasures translate directly into governance rules:
- Minimal tools. Only connect the extensions and functions the task needs. A summarizing agent does not need a send function.
- Least privilege in the user's context. Actions on downstream systems run with the rights of the specific user and the minimum scope, not with a high-privileged generic identity.
- Human approval for high-impact actions. Payments, external emails, deletions and permission changes wait for a person.
- Authorization in the target system. The downstream system decides whether an action is allowed, not the model.
- Logging and monitoring. Record what the agent did, so you can spot actions that should not happen and respond.
The OWASP Top 10 for Agentic Applications for 2026, published on December 9, 2025 by the OWASP GenAI Security Project, extends this to autonomous agents and uses the term "Least-Agency" explicitly. Three of its ten entries map straight onto governance work: ASI03 Identity and Privilege Abuse (agents inheriting or escalating credentials through delegation), ASI09 Human-Agent Trust Exploitation (people approving agent actions without independent checks) and ASI10 Rogue Agents (agents that drift from their intended function or authorized scope).
Identity is the open engineering question. In February 2026, NIST's National Cybersecurity Center of Excellence published a concept paper on identity and authorization for software and AI agents and asked for feedback on identification, authorization, auditing and non-repudiation of agents. It is a concept paper for a potential project, not a standard. The practical takeaway for now: give agents their own identities instead of shared service accounts or borrowed user tokens, so that every action can be traced to one agent and one owner.
Which AI agent governance frameworks help?
None of the frameworks below is an AI-agent law. They give you vocabulary and structure for the inventory, owner and permission rules above.
| Framework | What it gives you | Status |
|---|---|---|
| NIST AI Risk Management Framework 1.0 | Four functions (govern, map, measure, manage) for AI risk; Generative AI Profile NIST AI 600-1 from July 26, 2024 | Voluntary, released January 26, 2023 |
| OWASP LLM Top 10 2025 and Agentic Top 10 2026 | Concrete risk categories and mitigations for agents | Community project |
| Singapore Model AI Governance Framework for Agentic AI (IMDA) | Four dimensions: bound risks upfront, make humans accountable through approval checkpoints, technical controls across the agent lifecycle, end-user responsibility through transparency and training | Guidance, launched January 22, 2026; non-EU reference |
| NIST NCCoE concept paper on agent identity | Questions to ask about agent identification, authorization and audit | Concept paper, comment period closed April 2, 2026 |
For a broader overview of how these frameworks differ, see our glossary entry on AI risk management frameworks.
What does the EU AI Act require from companies that deploy AI agents?
If your company uses AI agents for business purposes, it is very likely a "deployer" under the AI Act and already has to take measures to support the AI literacy of the people who operate those agents. Stricter deployer duties apply only to high-risk systems, and most internal productivity agents will probably not be high-risk. Counsel should confirm both points for your setup.
Shadow agents are not exempt. The AI Act does not distinguish between approved and unapproved use. In our reading, when an employee uses an agent for work, the company is typically its deployer, whether or not IT knows about the agent. That is one more reason to find shadow agents first.
Who is a deployer. The Act defines a deployer as anyone "using an AI system under its authority" outside personal, non-professional use (Art. 3(4)). Switching on or building an agent for business use typically makes the company a deployer. If you develop an agent yourself and put it into service, you may also count as a provider; counsel should assess.
AI literacy (Art. 4). The duty applies since February 2, 2025, and national market surveillance authorities supervise and enforce it from August 2, 2026, according to the European Commission's AI literacy Q&A. The Digital Omnibus on AI, which entered into force on July 27, 2026 according to the Commission's AI Act overview, changed the wording: providers and deployers "shall take measures to support the development of AI literacy" of their staff, and the duty "does not require" them to guarantee any specific level. The obligation remains. The Commission says no certificate is needed and that an internal record of trainings is a way to document it. Its Q&A also says that even employees using ChatGPT for routine text work should be informed about specific risks such as hallucination. Agent owners and the people who approve agent actions are the obvious first audience.
High-risk duties (Art. 26). Deployers of high-risk AI systems must, among other things, assign human oversight to people "who have the necessary competence, training and authority" (Art. 26). After the Omnibus, the rules for high-risk use cases in sensitive areas, including employment and critical infrastructure, apply from December 2, 2027, and those for AI in regulated products from August 2, 2028. An agent that screens job applicants or allocates work could land in a high-risk area; have those classified first.
Agents that talk to people (Art. 50). AI systems that interact directly with people must be designed so those people know they are dealing with AI (Art. 50(1)). That is mainly a provider duty, relevant for customer-facing agents; the Commission dates the transparency rules to August 2026. Our article on deepfake simulations, consent and the AI Act covers Art. 50 in more detail.
GDPR still applies. When an agent processes personal data, the GDPR applies alongside the AI Act, including the question of whether a data protection impact assessment is needed.
The links to Articles 3, 4, 26 and 50 go to artificialintelligenceact.eu, an unofficial consolidated version maintained by the Future of Life Institute; the binding text is published in the EU's Official Journal. This is not legal advice.
The human side of agent governance
Every shadow agent starts with a person who wanted to get work done faster. That makes governance a people topic as much as an IAM topic, in three ways.
People build and connect agents. A ban alone tends to push agent use further out of sight. A clear policy that says which platforms are approved, how to register an agent and what owners are responsible for, combined with short training for the people who build agents, brings agents into the inventory instead.
Agents can be talked into things. An agent that reads emails, web pages or documents can receive instructions hidden in that content. Our blog post on how prompt injection hijacks AI agents explains the mechanism. Governance answers it with the permission limits from Step 3, not with model trust.
People trust agents too much. OWASP's ASI09 describes how users approve agent actions "without independent validation" when an agent sounds confident. An approval click becomes a control only if the approver knows what to check, the same way an employee learns to verify a payment change by calling a known number. Approvers need training on that, just as they do for phishing.
Example scenario (fictional): A supplier email contains hidden text telling the finance team's inbox agent to update the supplier's bank details. The agent prepares the change and asks an accounts payable clerk to approve it. The clerk does not rely on the agent's summary, calls the supplier on the number in the master data, learns that nothing changed, rejects the request and reports the email to security.
How revel8 fits
revel8 is not an agent inventory or AI governance tool, and this guide does not assume you use it. Two parts of the platform address the human side and customer-facing agents:
- Contact Center Protection pentests customer-service AI agents: it maps the channels the agents operate across, flags likely weak points and tests them with current agentic prompt-injection techniques. It also runs AI voice simulations against the human agents behind them.
- Interactive Trainings drafts modules from documents you upload, such as an internal AI or agent policy, so you can turn that policy into training built from your own policies for agent owners and approvers. revel8 describes reporting as works-council conform and without individual surveillance.
If you want to test your customer-facing AI agents and the people around them, you can book a demo.
A 30-day plan to bring shadow AI agents under control
You do not need a governance platform to start. A spreadsheet with the fields from Step 1 and four focused weeks get you from "no idea" to a first defensible state:
| Week | Goal | Result |
|---|---|---|
| 1 | Discover | Export OAuth grants and SaaS agent settings, review low-code admin views, launch the self-reporting amnesty |
| 2 | Assign owners | Every agent has a business owner, a technical owner and deputies; agents without an owner are suspended |
| 3 | Reduce rights | Unneeded tools and scopes removed, approval points set for high-impact actions, shared accounts replaced by agent identities where the platform allows |
| 4 | Make it routine | Agent policy published, AI-literacy measure for owners and approvers documented, review cycle and expiry dates set |
After day 30, the inventory review becomes part of your regular access reviews, and new agents go through registration before they get credentials. Two simple measures show whether governance holds: owner coverage (agents with a named business and technical owner ÷ all agents in the inventory) and review coverage (agents reviewed within their review cycle ÷ all agents in the inventory). Count suspended agents in the denominator until they are removed, otherwise suspending orphans makes the numbers look better than the situation is.




