← Stackzilla Blog

AI Tool Permissions: A Model Request Is Not User Authorization

Published November 16, 2026 · 3 min read · AI agents, security, authorization, tool calling

Put approval and access checks around AI tools so untrusted text cannot silently turn into permission to change accounts or send data.

An AI assistant that can summarize text has a different risk profile from one that can send email, change records, or initiate purchases. Tool access connects generated language to real effects. The model may propose an action, but the application must decide whether that action is allowed. Treating every tool call as authorized because the model produced valid JSON confuses syntax with permission. ## Keep authority outside the prompt Authenticate the user and enforce their permissions in the tool implementation. Do not let the model select an arbitrary tenant identifier or claim an administrator role in its arguments. Use narrow tools with explicit schemas. A tool that updates one permitted field is easier to secure than a general command runner with broad credentials. Validate values and resource ownership after parsing the model's request. Read access and write access should be separable. An assistant helping someone investigate an account issue may need to view records without having permission to delete or export them. ## Treat retrieved content as untrusted Web pages, documents, emails, and tool responses may contain instructions addressed to the assistant. Those instructions are data from another source, not permission from the user. Prompt injection can attempt to redirect the workflow, expose confidential information, or trigger actions. Clear prompt boundaries help, but they do not replace server-side controls. The safest design limits what an injected instruction could accomplish. Validate destinations for outbound messages and data transfers. A retrieved document naming an email address should not automatically authorize sending account data there. ## Require approval where consequences justify it For significant actions, show the user exactly what will happen before execution: the recipient, resource, amount, or irreversible change. Approval should bind to that specific operation, not grant broad permission for later actions. If the proposed arguments change after approval, require the appropriate review again. Prevent retries from repeating side effects by using idempotency or an equivalent domain safeguard. Record an audit trail that distinguishes the user's request, the model's proposal, the authorization decision, and the tool result. Avoid retaining unnecessary sensitive content in those logs. Test with adversarial documents and misleading tool output, not only cooperative prompts. Include a case where an otherwise useful document contains an instruction to bypass the workflow. An effective AI agent is not the one with the fewest restrictions. It is the one that can complete the intended task while keeping identity, permission, and user consent under the application's control. ## Further reading - [OWASP: LLM prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) - [OWASP: LLM excessive agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)

Read the full article on Stackzilla →