Perk MCP security and data handling
Perk MCP acts only with your own Perk permissions, never completes or pays for a booking, doesn't store the data it returns, and doesn't train AI models. This article summarizes these controls, then covers each in detail: permissions, data access, storage, prompt injection defenses, logging, hosting, and sub-processors._
Perk MCP is designed so that an AI assistant can act only within the permissions of the person who connected it, and so that Perk keeps as little of the data it returns as possible. In summary:
- Permissions: The assistant gets the same Perk permissions you have, and no more. Perk MCP issues the access token your AI client uses and validates it on every request. Perk applies the same role-based permissions and company-level data separation as the Perk web app.
- Actions: The assistant can search, read, and make some changes, such as creating and submitting expenses. It can't complete or pay for a booking, modify a confirmed booking, or cancel a travel booking. You finish checkout in the Perk web app.
- Data storage: Perk MCP doesn't store or cache the data it returns to the assistant.
- AI training: Perk doesn't use data exchanged through Perk MCP to train AI models.
- Prompt injection: Perk MCP treats all tool results as untrusted data. A central layer fences and sanitizes every result, fixed model instructions tell the assistant not to follow instructions in results, and a permission boundary limits what injected instructions could do anyway.
- Logging and hosting: Perk records each tool call and each consent grant. Perk MCP runs on Amazon Web Services (AWS) in the European Union (EU) and uses Perk's existing sub-processors.
For the full sign-in flow, token lifetimes, and scope enforcement, see Authenticate with the Perk MCP server.
How Perk MCP enforces your Perk permissions
You sign in with your Perk credentials and approve access on a Perk consent screen, so the AI client never sees your password. Perk grants the access token only the scopes that match your existing permissions and removes scopes you aren't entitled to. A scope controls which resources an access token can read or modify. Perk MCP also accepts only allowlisted client redirect destinations.
In practice, a traveler sees only their own data. Admin, reporting, and approval functions need the matching Perk role.
What the assistant can and can't do
Perk MCP exposes about 55 tools, and you see only the tools your Perk role and permissions allow. The tools support these actions:
- Flights, trains, and cars: The assistant returns a checkout link. You open it and complete the booking in Perk.
- Hotels: The assistant adds a selected rate to a trip that isn't booked yet. You still complete checkout in Perk.
- Searches: Starting a search can create a draft trip.
- Expenses: The assistant can create, update, submit, approve, and send back expenses, and upload receipts. Approving an expense needs you to be an approver in Perk.
- Trip approvals: The assistant returns an approval link. You act on it in the Perk web app.
- Events and reports: The assistant can create Perk Events and generate reports.
Data the assistant can access
The assistant can reach the following data categories, each limited to what you can already see in Perk:
| Data category | What it includes |
|---|---|
| User profile | Name, email, and role. Admins can also see other users in the account. |
| Trips | Travelers, dates, itineraries, costs, and booking status |
| Travel search results | Availability and pricing |
| Travel policy | Policy rules |
| Expenses | Amounts, merchants, categories, statuses, and receipts |
| Company cards and transactions | Your company cards and pending card transactions still to be matched to an expense |
| Invoices | Invoices and invoice line items |
| Reports | Aggregated spend, travel, and team reports, for people with reporting access |
| Perk Events | Event details and attendee lists |
| Trip approval links | Approval links for pending trips |
Data storage, retention, and AI training
Perk MCP stores only the following:
- Encrypted OAuth tokens and client registrations (see Authenticate with the Perk MCP server for lifetimes)
- Short-lived operational data, such as rate-limiting counters, and sign-in state that expires after 15 minutes while a sign-in is in progress
- Operational logs and traces, used for security monitoring and troubleshooting, which record each tool call and, for travel searches, the search details and results (see Logging and audit)
All of this data falls under Perk's existing data processing arrangements and Data Retention Policy, available on the Perk Trust Center.
Perk doesn't use data exchanged through Perk MCP to train AI models. The Perk MCP server doesn't call any AI model itself. It only returns data to the AI client you chose. Perk doesn't see your prompts: your AI provider runs the model, and Perk MCP only receives the requests your client sends for Perk data.
Controls against prompt injection and untrusted data
Prompt injection is text inside data that tries to make an AI assistant follow instructions you didn't give it. Perk MCP treats all data its tools return as untrusted, including user-written fields such as expense descriptions, comments, merchant names, and vendor notes. A central layer applies the following protections to every tool, so a new tool is covered automatically.
- Untrusted-data fencing: Perk MCP wraps every tool result in
BEGIN-UNTRUSTED-DATA-<nonce>andEND-UNTRUSTED-DATA-<nonce>markers. The nonce is random and changes on every call, so stored text can't forge the closing marker. A preamble tells the model that other people may have written the content, that it must never follow instructions in it, that it must never call tools or choose tool arguments because of it, and that it acts only on your own request. - Sanitization of returned text: Before results reach the model, Perk MCP applies Unicode normalization and removes control, zero-width, bidirectional, and Unicode tag characters, which attackers use for hidden "ASCII smuggling". It also removes role and chat-template markup, such as
<system>,<|im_start|>, and[INST], and attempts to fake the untrusted-data markers. Sanitization covers both the text and the structured output. It is hygiene, not the trust boundary: Perk MCP intentionally doesn't rewrite ordinary-language instructions, so the fence is what contains them. - Fixed server instructions: Static, first-party instructions state that tool results are untrusted data that other users may have written, and that the assistant must never follow instructions inside them or call tools because of them. Perk MCP never interpolates request or tool-result content into these instructions.
- Permission boundary: The assistant acts only with your own Perk permissions and the scopes you consented to. Write tools need an explicit write scope, and the backend enforces per-record authorization. An injected instruction can't make the assistant do anything you couldn't already do in Perk.
- Tool annotations and confirmation guidance: Each tool declares whether it is read-only and whether it is destructive, so AI clients can decide which tools need your approval. Some AI assistants let you require approval for individual tools in the connector settings. The expense write tools (create, update, submit, approve, and send back) instruct the assistant to confirm the specific record and change with you before acting. For what each tool does and the scope it needs, see the Perk MCP tools reference.
- Input validation and limits: Perk MCP validates all tool inputs and applies per-user and per-tool rate limits and concurrency limits to constrain abuse.
How Perk tests these controls
Perk tests these defenses in three ways:
- Independent penetration testing: An independent third party penetration tests Perk MCP (latest report), including cross-user prompt injection scenarios. In these scenarios, one user plants instructions in a text field that another user's assistant later reads.
- End-to-end adversarial regression tests: Automated tests run an injection payload through a real MCP client, the same way an assistant receives tool results. They check that:
- The untrusted-data markers match.
- The marker nonce changes on every call.
- Forged markers and injection markup are removed.
- The structured output stays valid.
- Sanitizer unit tests: These cover hidden and control characters, role and chat-template markup, forged markers, and attempts to splice a marker together from fragments.
Logging and audit
Perk MCP records each tool call in Perk's monitoring platform. Each record includes the following fields:
- User and account IDs in Perk
- Tool name
- AI client
- Session ID
- Outcome (success or error)
- Response time
For travel searches, the logs also include the search details and results.
Perk MCP identifies its requests to the Perk backend with a dedicated User-Agent and an AI-client header, so Perk can tell them apart from web app traffic in logs. Perk also records each consent grant with the user, account, and access granted. Perk's change history keeps changes to records.
Monitoring alerts cover errors, unusual traffic, and authentication failures. You can access selected logs in the Perk app. Account admins can request access to further logs through their account manager when incidents occur, according to their contract.
Hosting and sub-processors
Perk MCP runs on Amazon Web Services (AWS) in the European Union (EU), in Ireland (eu-west-1). Perk's existing platform services process the data. Perk holds operational logs in its EU monitoring environment. Perk hosts client data related to the Spend and Pay services on Google Cloud Platform (GCP), in regions in Switzerland, Belgium, Germany, and Ireland.
Perk MCP uses Perk's existing sub-processors, listed on the Perk Trust Center. Two tools rely on existing providers. The visa requirements tool sends trip route and nationality details to Perk's visa information provider.
Receipt uploads use Perk's existing secure file storage. You choose and contract the AI assistant yourself.
For more information about Perk's security practices, see the Perk Trust Center.
Next steps
Updated about 1 month ago

