Skip to content
ConsultEvo

AI Security Risks: Controls, Approvals, and Safe Automation

Manage AI security risks by inventorying tools and data flows, assigning each use case an owner and impact tier, applying controls that match the exposure, and requiring human approval for consequential or irreversible actions. A support-ticket summary may be screened and written to approved CRM fields after validation. A proposed payment-account change should be independently verified and approved before any system acts on it.

The goal is governed use, not a blanket ban. AI security risk is the possibility that AI use, AI-enabled applications, or AI-connected actions expose data, credentials, systems, or people to harm. This guide presents an operational design approach. No single product eliminates these risks.

In Gallup’s October 2024 survey, 70% of employees said their organization did not have guidance or policies for using AI at work. That is a dated survey finding, not a current universal estimate. Read Gallup’s finding and context.

How should an organization manage AI security risks?

Use four steps: inventory AI tools and data flows, classify potential impact, assign controls and an accountable owner, then monitor and revise. Create an approved path that meets the employee’s actual need, so safe use is practical rather than merely prohibited.

Automate only after the data, permission, decision owner, and exception path are defined.

For each use case, record what information enters the model, what the model can do, which accounts and destination fields are involved, and who handles a blocked, stale, failed, or uncertain run. A low-impact internal summary and an automated payment-detail change should not share the same approval policy.

Map seven common AI security risks to controls

Use the table to choose a first-line control and define the condition that requires escalation. It is a risk-orientation tool, not a complete security assessment.

Risk area First-line control Escalation trigger
Shadow AI Inventory tools, provide a fast approval path, and offer a sanctioned alternative. An unapproved service receives restricted data.
Data leaks Classify and minimize data; confirm service terms before submission. Restricted or regulated data is exposed or retained unexpectedly.
Credential theft Use strong authentication, key management, and periodic access reviews. An unusual sign-in, exposed key, or unnecessary access is identified.
Generated code vulnerabilities Review code, scan secrets and dependencies, and run security tests. A change affects authentication, authorization, payments, or sensitive data.
Prompt injection Treat input and retrieved text as untrusted and restrict available tools. Input requests a tool action outside the task or policy.
Insecure AI applications Review architecture, data handling, permissions, retention, and response plans. Access, retention, incident notice, or recovery arrangements are unclear.
Deepfake-enabled fraud Verify unusual high-impact requests through a separate channel. A request changes payment details, credentials, permissions, or creates an irreversible result.

These risks need different responses. A 2021 study generated 1,689 programs across 89 scenarios and reported vulnerable outputs in approximately 40% of its GitHub Copilot-era test cases. It supports reviewing and testing generated code, not applying that result as a current vulnerability rate for all coding tools. See the study and its scope.

Credential exposure also requires precise interpretation. IBM X-Force reported that more than 300,000 ChatGPT credential sets were advertised on the dark web in 2025. The figure concerns advertised credential sets, not a confirmed count of compromised accounts. Read IBM X-Force’s report.

Use deterministic rules for exact conditions and AI for context

Use ordinary validation or policy rules when the condition is explicit and auditable. Examples include approved domains, known secret formats, required fields, allowed CRM values, and whether an action requires approval. These checks are easier to test and explain than asking a model to infer whether an exact condition has been met.

AI classification can help with contextual questions, such as whether text appears to contain confidential strategy or resembles an attempt to manipulate an agent. Treat that result as a screening input, not proof that content is safe. Combine both approaches when useful: deterministic rules enforce hard boundaries, while AI flags ambiguous cases for routing or review.

Zapier documents AI Guardrails actions for PII, prompt attacks, toxicity, and sentiment. Its documentation warns that results can include false positives and false negatives and describes the product as supplemental to existing controls. Builders must configure what happens next, such as filtering, blocking, routing, or requesting review. Detection alone does not guarantee those outcomes. Review AI Guardrails setup, scope, and limitations.

For example, a deterministic pattern can block a known API-key format before text is sent onward. A semantic detector can flag a message telling an agent to ignore its instructions. If the result is ambiguous, route it to a named security or workflow owner. Define timeout and unexpected-output behavior. For sensitive actions, hold the run for review rather than treating a missing result as approval.

Decision point

Use a rule for a hard boundary and semantic screening for uncertainty. A known secret pattern can be blocked deterministically. A possible prompt attack should be assessed in context and routed to a named reviewer when the outcome is uncertain.

Design an AI workflow with a real decision gate

A model result should not flow directly into a system of record. The following is a proposed CRM design, not a claim that one vendor feature supplies the complete workflow. It starts with a changed support or CRM record, sends only permitted text for a bounded classification, validates the response and source version, applies policy checks, and writes only approved fields.

01Capture the source eventKeep the source system, record ID, version or updated timestamp, and original event. The CRM or ticketing system remains the system of record.
02Minimize input and bound the taskSend only permitted text. Request a defined classification or summary, not an open-ended action plan or authorization decision.
03Parse and validate the resultCheck required fields, allowed labels, source identity, record freshness, confidence handling, and field-level permissions before any write.
04Apply policy and the decision gateUse deterministic rules and configured screening. A named reviewer handles uncertain results, high-impact fields, and consequential changes.
05Write approved fields and retain provenanceUpdate only permitted fields, record the run and decision, and send failures, stale records, declines, and timeouts to an exception owner.

A proposed output contract, for illustration only, is:

{
  "source_record_id": "ticket-illustrative-1042",
  "source_updated_at": "2026-11-20T10:15:00Z",
  "classification": "billing_question",
  "confidence": 0.91,
  "screening_decision": "review",
  "review_status": "pending",
  "model_version": "example-model-version",
  "prompt_version": "support-triage-v3"
}

Allow only explicitly defined classification values. Verify that the source record has not changed since the trigger. Check that the connected account can write only the intended fields. If a reviewer declines or does not complete review, do not write the proposed classification.

Zapier documents a Human in the Loop approval action that can pause a Zap and return a decision for downstream logic. Independent verification and the specific CRM policy remain organizational design choices. See the documented approval action. For related process design, see CRM workflow and data design.

Require human approval for consequential actions

Use a separate decision path for payment changes, permission changes, sensitive exports, destructive operations, legal-status updates, and other actions that are difficult to reverse. AI may summarize a request or extract proposed fields, but it should not authorize the action or serve as the sole evidence for approval.

The approval record should retain the request ID, source record, proposed action, supporting evidence, reviewer identity, decision, and decision time. For unusual financial or executive requests, independently verify the request through a separate channel. Define expiry, decline, skipped-review, and reviewer-unavailable paths before enabling the action.

A low-risk internal summary might need validation and no approval. A customer-facing draft may need one reviewer. A payment destination change may need separate-channel verification and two-person approval. These are proposed risk tiers, not an official vendor template.

Scope tool access before an AI agent can act

Connecting an assistant to a business app does not itself make its actions safe. The connected account’s permissions determine what data and actions are available. OAuth or centralized credential handling does not make an overprivileged account least-privilege. Review app connection behavior.

For MCP-style access, inventory each enabled tool, its app and connected account, whether it reads or writes, its data scope, approval level, owner, and last review date. These are proposed governance fields, not a vendor schema. Where supported, use a fixed, narrow tool set for sensitive tasks rather than assuming discovery or a connection platform prevents unauthorized actions. Treat a remote or third-party MCP server as a separate trust boundary that needs review.

Zapier documents MCP as exposing selected app actions through connections it manages. Its Enterprise documentation describes managed apps and managed connections, among other administrative controls. These are plan-specific controls. Source-application permissions still need review. Read the MCP documentation and check the Enterprise connection controls. For bounded agent workflow planning, see AI agent design and implementation.

Review vendors, retention, and evidence before approval

Approve a service for a defined use case and data class only when its documented terms, access scope, accountable owner, and exception path are acceptable. Check the exact product and plan, not just a general vendor statement. Confirm retention, deletion, model-training use, staff access, region, subprocessors, incident notification, audit evidence, recovery, and support against contractual requirements.

Approval checks for the specific use case
  • Is the intended data class permitted for this exact service and plan?
  • Are retention, deletion, and model-training terms documented and acceptable?
  • Does the connected account have only the access the workflow needs?
  • Can the organization review relevant access, run, and incident evidence?
  • Is a named owner responsible for exceptions and access reviews?
  • Is there a defined response for exposure, outage, or unexpected processing?

Retention is part of workflow design. Zapier’s AI Guardrails documentation says processed data may be temporarily stored in logs and Zap runs. The documented periods vary by settings and plan, including seven days in logs and 29 to 69 days in Zap runs. Recheck current product documentation before approval, especially when screening sensitive text. Decide whether the workflow should retain raw content, redacted text, or only a reference and decision.

Prebuilt integrations may reduce the amount of authentication and maintenance code an organization operates, but they are not categorically safer than custom integrations. Compare the actual provider, permissions, configuration, updates, and monitoring. For workflow gates and exception handling, see Zapier workflow design.

Measure whether controls are working

Track approved-tool coverage, access-review completion, screening escalations, reviewer turnaround, rejected or corrected actions, and incidents. Tool usage and policy adoption describe activity; neither alone proves that security risk has fallen. Use incident reviews to adjust scopes, thresholds, approval rules, and employee guidance.

Keep each source event, AI run, detection, citation, and aggregate at its own data grain. One run record represents one model execution and retains its source event and execution identity. Each detection or citation is a separate observation when there are multiple findings. Daily or monthly metrics belong in a separate aggregate record with an explicit period, model or detector, and segment definition.

A proposed event record could include source_system, source_event_id, source_version, run_id, model_version, prompt_version, decision, reviewer_outcome, destination_record_id, and execution_status. These are implementation suggestions, not a vendor-published schema.

Use identifiers that distinguish retries and independent runs. For a source event, a proposed key is source_system + source_event_id + source_version. For a model run, add model_version + prompt_version + execution_uuid. A finding should reference its run and detector type. A daily aggregate should include the metric, period, engine or model variant, and segment.

Where the destination supports it, enforce uniqueness with a database constraint and use an atomic or transactional upsert. A lookup followed by create can race when concurrent runs process the same event. If atomic upsert is unavailable, use a durable idempotency key, locking, or a single-writer queue. Keep CRM contact or deal events separate from raw AI observations and reported summary metrics.

Recheck vendor documentation, feature availability, retention, and plan restrictions immediately before publication or rollout because product details can change.