# AI Security Risk Assessment

> Map AI data access and tool authority before widening a pilot. Prioritize unresolved workflow boundaries with qualified owners, not a self-certification score.

Source: https://grandrounds.tech/insights/ai-security-risk-assessment · Published 2026-11-13 · CTO Grand Rounds

You cannot establish the workflow's reach from its interface or model instructions. Before widening a pilot, have qualified owners inspect evidence at each data and authority crossing, starting with the most consequential unresolved exposure. If you need an AI risk assessment tool, this page offers a worksheet, not software or certification; a product purchase depends on which evidence gap remains.

*Figure: A draft document sits within a dotted task boundary while a separate sending port is connected through a question mark, asking whether action reach exceeds the task.*

**Can the AI reach data or tools that the user should never control?** Before widening a customer pilot, answer that question across the workflow—not only at login or in the model's instructions.

For a software company moving an AI prototype into use by a named customer, the practical output is a risk-evidence record: permitted data and actions, the mechanisms that enforce those boundaries, what has actually been reviewed, and who owns unresolved questions. This article helps prepare that decision. It is not a penetration-test guide, a security certification, or a substitute for qualified security review.

## Scope one customer task and its authority

Choose a bounded journey, such as preparing a support-response draft from a customer's approved knowledge base for a human reviewer. Record the user group, tenant, source collection, output destination, deployment version, and excluded operations. Sending a reply, changing a customer record, or issuing a refund is a separate capability—not permission inherited from drafting.

Identify the actual technical, data, security and domain decision holders, including the customer's acceptance requirements. Confirm who can narrow the customer promise and resource the review; a senior title does not cover every responsibility.

Keep the existing manual route as a credible option. A founder manually rescuing the demo is not a long-term control, but the team may be able to narrow the pilot to a supported draft-only task while an unresolved boundary is reviewed.

## Follow data and authority through the workflow

Make a short map of what enters, what is retrieved, what reaches the model, which tools can be called, and where outputs or actions go. Include saved results, logs, exports, support access, and background jobs—not only the visible chat.

At each crossing, ask:

1. **Whose identity and authority apply?** The end user, a service account, a background worker, or another system?
2. **What may cross?** Which data, tenant, purpose, and operation are permitted?
3. **What enforces that rule?** Where is the check, and what evidence can a qualified reviewer inspect?
4. **What happens when the rule cannot be established?** Who receives a denial, exception, or uncertain action outcome?

Separate data as evidence from data as instructions. [OWASP's LLM01:2025 guidance](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) describes indirect prompt injection through external sources such as files and websites, and notes that RAG does not fully mitigate prompt injection. A retrieved document is not a grant of new authority merely because it was relevant to the question.

*Figure: Map input identity, permitted sources, model context, actual tool reach and destination. Include saved results, logs, exports, support access and background jobs. Name an owner at each crossing; the map proves no control. — Lines locate review questions; they do not establish authorized flows or passed controls.*

## Use the AI workflow risk checklist

For each row, record the **current claim, evidence reference/version/date, reviewer and remit, finding or unknown, owner, next bounded review, and status**. Distinguish documented, inspected, observed, and accepted. Start missing evidence as **Unknown**, not “secure” or “failed.”

Where AI can call tools or propose actions—the concern often called **agentic AI security**—the Tool reach and External action rows help locate the additional exposure. They do not replace the access and data rows.

| Boundary | Evidence questions for the responsible owners | What not to infer |
| --- | --- | --- |
| **Data use and storage** | Which input/source types are permitted for this purpose? Where do prompts, retrieved context, outputs, caches, and logs go? What retention, deletion, provider, and support-access conditions apply? | Possessing a file or having a vendor account does not establish rights for every use. |
| **User, role, and tenant** | Which identity controls retrieval, saved answers, exports, and background work? What evidence covers permitted and denied paths and changed/revoked access? | Successful login is not proof of document or tenant isolation. |
| **Untrusted content** | Which external/user-controlled inputs can influence output or tool selection? Which independent boundaries limit the consequences? What remains unexamined? | A system prompt, citation, or content filter is not sufficient evidence of authorization. |
| **Tool reach** | Which operations exist, which are actually needed, and whose downstream permissions apply? Are read, write, delete, send, and administrative functions distinguished? | A tool's friendly name does not establish its real authority. |
| **External action** | What records exact-action authority and shows execution waits for that decision? | A model's proposal, a UI approval label, or a generic signed-in user is not necessarily authorization. |
| **Failure and uncertain outcome** | Who reconciles attempts with destination results after timeouts, repetition, rejection or changed evidence? | No confirmation is not proof that nothing happened; approval does not establish successful execution. |
| **Operation and escalation** | Who can restrict the workflow, preserve necessary evidence, notify the relevant owners, and restore an agreed fallback? Are support, reviewer capacity, dependency changes, and incident responsibilities established? | Monitoring alone does not prove prevention or an effective response. |

This checklist is a discussion aid, not an exhaustive threat model or self-approval rule. Your security specialist determines the appropriate assessment and assurance work for the actual architecture, data, exposure, and consequences. Privacy, legal, contractual, and domain decisions retain their own qualified owners.

For approval-packet and recovery design, see Human-in-the-Loop AI Agents: Who Can Approve External Actions? A customer questionnaire is a different evidence task, covered in Enterprise SaaS Readiness: Turn a Security Questionnaire Into Evidence.

## Narrow the pilot first: what each route adds to the boundary

[OWASP's LLM06:2025 guidance on excessive agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) identifies excessive functionality, permissions, and autonomy as root causes, and recommends downstream authorization rather than leaving permission decisions to the model. Use that as a review prompt—not a claim that a particular configuration is safe.

For the same task of preparing a reviewed support response, compare:

| Route | What changes in the boundary discussion | Evidence still needed |
| --- | --- | --- |
| Existing search/templates and human preparation/release | No model-directed tool action is needed for this task | Source permissions, professional review, confidentiality, and ordinary application controls |
| AI drafts from permitted sources; human releases through the existing channel | Drafting and release remain separate | Retrieval/context/output access, evidence support, handling of untrusted input, reviewer authority and capacity |
| AI workflow proposes a channel action behind an authorization step | Adds an action path and more integration/recovery questions | All draft-path evidence plus action-specific authority, gating, destination result, and exception handling |

Compare actual coverage, effort, and remaining responsibility on one journey. A narrower route can reduce the capability being evaluated without proving the rest of the application secure. A new governance platform is not automatically necessary; your current stack may support the required boundary after an appropriate review.

## Invented example: a draft task with a wider connector

*Figure: Invented draft-only pilot has a connector inventory listing article reading and message sending. Configured permissions and model reach are unknown. The finding is action reach unverified against the promise, not a breach; imported instructions confer no authority.*

**Invented scenario—no customer data, system, attack or test.** A hypothetical pilot promises draft support replies. Its design inventory lists a connector that can both read knowledge articles and send messages. The intended user needs drafts, not sending authority. The inventory does not establish the connector's configured permissions or whether the model can reach the send operation.

The right finding is **“action reach unverified against a draft-only promise,”** not “a breach occurred.” The next bounded review is to have the authorized technical/security owners inspect actual tool exposure, downstream identity and permissions, and the independent execution boundary.

Add a synthetic discussion variant in which an imported document contains a request to change the response destination. No payload or action is executed. The review question is whether source content could influence an operation beyond the user's mandate, and which boundary would prevent that. A more persuasive answer or a stronger drafting model would not itself establish that boundary.

The team could propose retaining the manual release path, limiting the pilot's action scope, or evaluating an action-specific authorization design. Those are options for accountable owners to accept—not instructions here to modify live permissions. If sending later becomes a genuine requirement, its authority, evidence, and recovery scope need separate review.

## Turn the checklist into a bounded decision

*Figure: Blank record asks for unresolved crossing, explicit scope and authorization reference, appropriate environment, permitted fixtures, qualified oversight and evidence needed. Unconfirmed fields remain unknown. A plan is not permission to test. — A plan is not permission to test.*

Complete the checklist with the people who own data, technical operation, security, and output use. Choose the first material unresolved crossing and specify the evidence that would resolve it. Record whether the decision is to keep the current route, narrow the pilot, arrange specialist assessment, or plan a bounded repair and authorized verification.

Prioritize by what could be disclosed, changed or sent, to whom, and how easily it could be triggered. Your security specialist sets the final order; this is discussion advice, not a validated risk score. For example, an unresolved cross-tenant read may be more consequential than sending a draft back to the same authorized user. “Sending exists” is not by itself a reason to rank that row first.

Any testing requires explicit scope and authorization, an appropriate environment, permitted fixtures, and qualified oversight. This page provides no exploit payloads or instructions to probe production. Do not expose live confidential records to discover whether a control works. If an actual suspected incident arises, use your established incident route rather than improvising an investigation from this checklist.

Acceptance should name the version, user group, permitted sources/actions, reviewers who actually participated, unresolved limits, receiving owner, and changes that reopen review. An unchecked row does not disappear under an overall score.

If the unresolved question spans product scope, access, tools and operating ownership, [ask the Grand Rounds curator about the relevant next conversation](https://grandrounds.tech/talk-with-the-curator). Specialist security assessment and release remain with your qualified owners; this is not an offer of assessment or implementation.

## Boundaries

Grand Rounds is an initiative of Dinarys, and a conversation may lead to cooperation with Dinarys.
