In a vendor-sponsored Cloud Security Alliance survey conducted in January 2026, 82% of 418 IT and security respondents reported discovering previously unknown AI agents during the previous year. That finding makes approval validation a practical security task: teams need to establish who owns a discovered agent and whether its access is authorized. CSA’s April 2026 survey release describes the sample and methodology; the result is self-reported, and unknown agents are not necessarily unauthorized.
Shadow AI agents are AI systems used outside an organization’s required approval or oversight processes that can select or execute actions toward a goal. They may query enterprise data, invoke tools, modify records, or communicate with external services through access that security teams have not reviewed.
Managing them requires more than a list of AI applications. Discover the deployment, identify the authority it uses, compare approved and actual scope, and decide whether to approve, restrict, or retire it. Verify that the chosen action changes what the agent can do.
What are shadow AI agents?
An agent becomes shadow AI when its deployment or use bypasses the approval and oversight required for that activity. An approved AI platform can still host an unapproved agent, and an approved agent can acquire tools or permissions outside its reviewed scope.
Separate unknown, unsanctioned, and compromised agents
These labels answer different questions. Keep approval status separate from evidence of excessive access or malicious activity.
| Finding |
What it establishes |
Appropriate next step |
| Unknown agent |
Ownership, purpose, or approval is unresolved |
Identify the deployment and validate its records |
| Unsanctioned agent |
Required approval or oversight is missing |
Assess exposure and choose a governed disposition |
| Overprivileged agent |
Granted access exceeds the task’s requirements |
Reduce permissions and review prior activity |
| Compromised agent |
Evidence indicates attacker control or misuse |
Follow incident response and verify containment |
A newly discovered deployment may be legitimate but missing from central records. Conversely, a registered agent may be compromised. Treat discovery as the start of a decision, and document which findings the evidence supports.
Distinguish agents from other AI use
A person pasting information into an unapproved chatbot is using shadow AI. An agent that retrieves files, selects a tool, and sends a message introduces an additional execution path. Its effective authority depends on its configuration, identities, connected tools, and downstream access.
For this guide, an agent uses an AI model to select or perform steps toward a goal. A model endpoint, connector, or Model Context Protocol (MCP) server can support an agent without being an independent agent itself. Record these dependencies because they can extend what the deployment can reach.
Why shadow agents need a different response?
Agents can keep acting through granted credentials, scheduled tasks, and connected services after the interaction that started a workflow. Reviewing the employee’s AI use therefore needs to extend to the actions and access of the resulting automation.
Adoption increases the importance of that review. In its August 2026 global survey, McKinsey reported that 40% of respondents from organizations with more than USD 1 billion in annual revenue were scaling AI agents, up from 27% in the previous survey. The study surveyed 1,719 participants across 97 nations between May and June 2026. This is a self-reported adoption comparison, not evidence that shadow-agent incidents rose at the same rate. McKinsey’s State of AI research provides the scope and methodology.
Shadow deployments can take several forms:
- A sales team connects an agent to a customer database using a shared account.
- A developer gives a local coding agent repository access and shell execution.
- A software as a service (SaaS) administrator enables an embedded agent without reviewing its data permissions.
- A temporary automation remains scheduled after its business project ends.
These are illustrative patterns. For each, the operational question is the same: what can act, with whose authority, against which resources, and under whose responsibility?
What risks do shadow AI agents create?
Risk depends on the agent’s access and ability to act, the sensitivity of connected resources, and the controls around consequential operations. Missing approval is a governance finding; sensitive reach or harmful activity determines how urgently the team should respond.
Delegated access can exceed the task
An agent may use a dedicated workload identity, an application identity, a service account, or a user’s delegated access. A narrow task can inherit broader privileges when a connector uses an existing credential.
For example, an agent asked to summarize selected documents might receive access to an entire shared repository. Assess the resources and actions actually granted, including access through downstream tools. Non-human identity security provides broader context for managing these credentials.
OAuth, a framework for delegated authorization, does not by itself establish that an agent’s business use is approved. The IETF’s OAuth security best current practice addresses restrictions on token privileges and protection of refresh tokens. Apply those controls to the actual authorization relationship.
Untrusted content can influence tool use
An agent may encounter instructions inside email, documents, web pages, or tool responses. If it treats that content as authority, a prompt injection can influence subsequent actions. The consequence depends on which tools and permissions the application allows.
OWASP’s Q1 2026 exploit roundup describes selected incidents and research disclosures involving tool misuse, identity abuse, and integration weaknesses. It is a non-exhaustive technical review, rather than a measure of how often all enterprises experience these events.
MCP introduces its own authorization boundaries. Its versioned security guidance explains why accepting tokens intended for other services and passing them downstream can undermine access controls and auditability. Approving an MCP server’s name or address is insufficient; review the identities, tools, and resources behind it.
For broader hardening requirements, see AI agent security.
Forgotten deployments retain authority
A deployment can stop serving a useful purpose while its schedules, credentials, or connectors remain active. Deleting a chat or removing an interface may leave other execution paths intact.
Retirement deserves its own review. NIST’s AI Risk Management Framework Playbook addresses inventory accountability in GOVERN 1.6 and deliberate decommissioning in GOVERN 1.7, including dependencies and evidence retention. Record what must be disabled, who can do it, and how the team will confirm the result.
How to discover shadow AI agents?
Combine records of what has been deployed with evidence of identities, tool calls, and activity. Each source reveals part of the picture; none should be treated as a complete inventory without a defined scope and verified coverage.
| Evidence source |
What it can reveal |
Important limitation |
| Platform and deployment records |
Agent configurations, owners, tools, and schedules |
Misses unsupported platforms and unregistered local deployments |
| Identity and authorization records |
Application identities, grants, service accounts, and access changes |
A shared identity may serve several workloads |
| Endpoint and development evidence |
Local runtimes, configurations, packages, and execution |
Installed software does not prove active agent use |
| Application and tool-call logs |
Actions, resources, run identifiers, and outcomes |
Coverage depends on instrumentation and retention |
| Network and service-access metadata |
Destinations, timing, communication paths, and data movement patterns |
Traffic alone may not establish agent identity or reveal encrypted content |
Use approved, read-only discovery methods. Do not run an unfamiliar agent or invoke an unknown tool merely to determine what it does. Record unsupported environments, collection failures, and the last successful collection time.
Correlate deployments, identities, and actions
Associate a finding with a stable deployment identifier, tenant or environment, owner, runtime, and credential references. Preserve the relationship between the initiating user, agent run, tool invocation, and downstream action where telemetry supports it.
A connection to an AI service is a lead. It could represent a person using a chatbot, a conventional application, or an agent. Likewise, activity under a shared service account cannot reliably identify which agent acted without additional context.
Build on a maintained AI agent inventory, and keep unresolved attribution visible. Store references to credentials rather than copying secret values into the inventory.
Validate approval against the current deployment
Check approval evidence against the configuration that produced the activity. An old approval for read-only testing does not automatically authorize production writes or a new external connector.
Confirm the accountable owner, business purpose, allowed resources and actions, permitted destinations, required human approvals, and review or expiry date. Record the effective configuration and decision time so investigators can distinguish a later permission change from the conditions present during a run.
How to prioritize a discovered agent?
Compare three things: what the organization approved, what the deployment was granted, and what the evidence shows it did. Differences identify separate problems: missing approval, excessive access, or activity outside expected behavior.
A useful decision record connects the agent to:
- Its business purpose and accountable owner.
- The identities and delegated users it can act through.
- The resources it can reach and operations it can perform.
- The external destinations and tools it may use.
- The actions that require human authorization.
- A review date and a verified method of stopping execution or access.
Use the following proposed decision aid to select the next action. Apply your incident severity process and business dependencies when setting urgency.
| Observed condition |
Priority consideration |
Suggested action |
| Active harmful activity or credible compromise |
Ongoing impact and reachable systems |
Start incident response and contain the relevant execution or access path |
| Sensitive access or consequential actions with unresolved approval |
Exposure exists before ownership is settled |
Restrict high-risk operations while validating authority and purpose |
| Confirmed unapproved use with limited exposure and a legitimate need |
A governed replacement may preserve useful work |
Set a review deadline and onboard, migrate, or discontinue |
| Unused or expired deployment with retained access |
Authority persists without a current purpose |
Retire it, check dependencies, and verify access removal |
An approval record does not resolve an active compromise. Equally, an unfamiliar application name alone is insufficient evidence to declare an incident.
What to do when you find a shadow AI agent?
Choose a response that addresses the supported finding and assigns responsibility for the result. A governance violation may need a controlled migration, while an actively misused agent may require immediate incident handling.
Preserve evidence and contain harmful activity
Where feasible, preserve relevant configuration, approval records, identity events, tool-call logs, and affected-resource evidence without delaying urgent containment. Follow existing retention and incident procedures, and protect sensitive content captured in logs.
Identify the controls that can interrupt the actual path: a runtime, schedule, connector, application permission, credential, or downstream resource policy. Check shared dependencies before disabling an identity used by other services.
Keep the response record precise. An analyst recommendation, an issued command, and confirmed containment are different states. Record the action taken, its effective time, and evidence that the prohibited operation can no longer succeed.
Do not assume that changing a user password or deleting an application entry invalidates every token. Access tokens, refresh tokens, grants, and active sessions may require different handling. Verify behavior in the relevant platform and downstream service.
Approve, restrict, or retire the deployment
For useful automation that can operate safely, assign an owner and migrate it into an approved environment or control process. Narrow resource access, replace shared credentials where feasible, document allowed actions, and test the required approval boundaries before restoring consequential operations.
Where the full workflow cannot be approved, consider a constrained mode, such as drafting without sending or reading a limited dataset without modifying records. Confirm that the restriction is technically enforced.
For retirement, remove obsolete triggers, schedules, integrations, and agent-specific access. Preserve required evidence, review residual data, and confirm that the deployment cannot restart through a second path. Close the record only when the responsible owner can demonstrate the intended outcome.
Example: an unapproved sales agent with shared-file access
Consider an illustrative agent built to prepare account summaries. A salesperson connects it to shared documents and email through delegated access, then adds a tool that can send requests to an external service. This is a hypothetical workflow, not a reported customer incident.
The organization discovers the agent after an unusual access pattern. Investigation establishes that its document access is broader than the sales task requires and that no approval covers its external tool.
Suppose a retrieved document includes instructions to collect unrelated files and send them outside the organization. The team should investigate the complete sequence: the document the agent read, its subsequent tool request, the identity used, the destination, and whether any transfer succeeded.
Different evidence supports different conclusions. A blocked tool call can show an attempted action without proving successful exfiltration. A downstream log may show a completed request without establishing exactly which content it contained. Preserve those limits in the incident record.
The immediate response might suspend the external tool and narrow the file-access path. A governed replacement could retain the useful summarization function with a restricted document set and human approval for sending. Before restoring it, test that an unapproved destination is blocked and that the agent cannot use another credential or connector to perform the same action.
The scenario illustrates the practical value of linking an agent’s purpose, permissions, inputs, tool actions, and enforcement results.
How to prevent shadow agents from returning?
Make the approved route accessible enough for teams to use, and define controls around the work agents actually perform. A policy should tell builders how to obtain approval and what changes require another review.
Provide a clear intake process, accountable reviewers, supported deployment options, and reusable identity and logging patterns. Ask teams to declare data sources, tools, destinations, and consequential actions before production use. Give experiments an owner, limited access, and an expiry decision.
Require review when an agent gains a new identity, broader data access, a new tool, an external destination, or greater autonomy. Check whether SaaS feature changes alter the approved boundary. Keep time-limited exceptions visible and make ownership transfers explicit.
Human approval should apply to the specific consequential action with enough context to evaluate it. A generic confirmation at workflow startup may not cover a later payment, bulk export, or permission change. Enforce authorization in the application and tools rather than relying only on instructions to the model.
OWASP’s November 2026 Securing Agentic Applications V2 overview identifies identity, runtime protection, supply-chain security, and assurance testing as parts of lifecycle security. For the wider workflow discipline, see agentic AI security.
A practical shadow-agent control plan
Start with a declared environment and a small set of high-impact workflows, then expand as the evidence improves. This proposed schedule is an implementation aid; it is not a benchmark or a required remediation deadline.
| Period |
Main work |
Evidence of completion |
| Days one through seven |
Set scope and ownership; gather platform, identity, endpoint, and activity evidence |
Source coverage, known gaps, and unresolved findings recorded |
| Days eight through 14 |
Validate owners and approvals; examine sensitive access and external tools |
Each priority finding has a disposition, owner, and review date |
| Days 15 through 21 |
Restrict risky access; migrate useful workflows; retire obsolete deployments |
Recorded changes and checks of their effective behavior |
| Days 22 through 30 |
Exercise stopping controls; reconcile new activity; improve the intake process |
Test results, remaining gaps, and the next review assigned |
For an active incident, follow incident-response urgency rather than waiting for this schedule. For complex environments, retain the same evidence requirements while adjusting the rollout period to dependencies and available resources.
How to measure whether controls work?
Measure accountability and effective control within a stated scope. A falling count of discoveries can reflect improved governance, reduced collection coverage, or a broken collector, so interpret it alongside evidence quality.
Useful measures include:
- Ownership coverage: the share of discovered deployments with a validated accountable owner.
- Approval reconciliation: findings whose current configuration has a documented approval or disposition.
- Access exceptions: unresolved differences between approved and granted resources or actions, with age and sensitivity.
- Attribution quality: sampled consequential actions that can be traced to a deployment and responsible identity.
- Response verification: tested restrictions or retirement actions that achieved the intended result.
- Collection health: evidence-source freshness, failed collections, unsupported environments, and unresolved attribution.
Keep denominators explicit. Coverage of known deployments is not proof of coverage of every agent in the enterprise. Set targets after establishing a baseline, and retain the evidence behind closed findings.
Connecting shadow-agent governance to security operations
Governance records help a security operations center understand whose agent is involved and what it should be doing. Observed activity adds context about which identities, services, and resources it actually touched.
Vectra AI’s AI-native security and observability approach is relevant where those activities cross connected environments. Its asset and identity visibility use case describes visibility across network, identity, cloud, and SaaS, including relationships between communication paths, privileges, and exposure.
When evaluating that connection, ask which agent deployments can be attributed, which telemetry is available, what remains outside coverage, and which enforcement systems can carry out a response. Require evidence for the environment you operate; a visibility description alone does not establish universal agent discovery or containment.
If you are evaluating this workflow, request a Vectra AI demo focused on your agents’ identities, connected environments, and investigation requirements.
Note: This guide draws on public research and technical guidance. The scenario and implementation plan are proposed practices.