> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wherobots.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Waldo security and privacy

> Which data Waldo can access, where prompts and results are processed and stored, how long Waldo keeps them, and how Waldo handles AI model training.

Waldo is the AI agent built into Wherobots Cloud. It runs in the Wherobots control plane on AWS and calls AI models through Amazon Bedrock in the Wherobots AWS account. This page covers the same commitments as the AI and Waldo section of the [Wherobots security page](https://wherobots.com/security/).

> [!NOTE]
>
> **Waldo is in public preview.** It's available in every Organization Edition, including the free Professional Edition trial. Features and limits may change before general availability.

## Commitments

* **Your data doesn't train models.** Wherobots doesn't train or fine-tune AI models on your prompts, chat history, query results, or files. Amazon Bedrock doesn't store or use prompts and responses to train models.
* **Waldo works only with your permissions.** Every query and tool call that Waldo makes runs with the signed-in user's own authorization. Waldo can't reach data or take actions that the user couldn't.
* **Your Organization stays separate.** Waldo chats are tied to your Organization and users. Other Organizations can't see them, and Waldo doesn't use one customer's chats to answer another customer.
* **You approve changes.** Waldo asks before it writes or deletes data, updates a file, or creates a job, task, or schedule. Each approval covers one action with the exact arguments shown to you, unless you choose **Always allow** for that tool.
* **You review the results.** AI output can be wrong. Waldo shows the SQL it runs and the tools it calls, so you can check its work before you rely on it.

## What data Waldo can use

Waldo can read only the catalogs, tables, and files the signed-in user can access in Wherobots. Turning on Waldo doesn't give anyone access to new data.

| Source | Examples | Who can use it |
| - | - | - |
| Your Organization's catalogs | Your tables and storage integrations | Users with access in Wherobots |
| Wherobots Open Data | Overture Maps, US Census and TIGER, USGS NHD, Copernicus DEM, Meta canopy height, Sentinel-1 and Sentinel-2 | Everyone |
| Wherobots Pro Data | Additional curated datasets | Organizations with a Pro Data entitlement |
| My Files | Files you or Waldo saved | You |
| The public web | Web search excerpts and their sources | Organizations with web search on |

* **Chats are private.** Only the user who starts a chat can see it.
* **Page context is optional.** When you open Waldo from another page, that page is sent with your message. Click the **X** next to it to leave it out.

## What goes to the model

Your messages, the system prompt, and the results of the tools Waldo calls are sent to the model to produce each answer. Tool results include table names, schemas, and samples of query results.

Waldo uses models from OpenAI and Anthropic, served by Amazon Bedrock in the Wherobots AWS account. Requests don't go to the model makers' own APIs. The default model uses a Bedrock global inference profile, so a prompt can be processed in any AWS commercial region. Waldo data is stored only in the places listed in [Where Waldo data is stored](#where-waldo-data-is-stored).

### Web search

With web search on, Waldo sends search queries to the Amazon Bedrock AgentCore Web Search API from the Wherobots AWS account. Waldo's instructions say to write search queries in plain words and never to include your data, credentials, or table contents. Admins can turn web search off in Organization Settings.

## Safety controls

* **Content screening.** Amazon Bedrock Guardrails checks each message you send for prompt-attack and harmful-content patterns before the message reaches the model. If the check can't run, Waldo stops the request instead of sending the message unscreened.
* **Prompt injection.** Data that Waldo reads can contain text written to mislead an AI model. Waldo acts only with your permissions, actions that change data need your approval, and the links Waldo creates point only to wherobots.com and its subdomains. With web search on, answers can also name the public pages they cite.
* **Limits.** Each Organization has a token quota, and each request has limits on time and steps, so a runaway request stops on its own.

## Where Waldo data is stored

| Data | Where it's stored | How long it's kept |
| - | - | - |
| Chats: messages, tool calls and results, saved maps | Wherobots control plane database (Amazon Aurora, encrypted) | Until the user or Organization is deleted, then 30 days in backups |
| Query result files | Wherobots temporary storage (Amazon S3) | 7 days |
| AI traces used for quality review: full chat content | Wherobots-hosted tracing service in the Wherobots AWS account | 60 days, then 30 days in backups, 90 days in all |
| Workflow state | Temporal Cloud | IDs only, payloads encrypted, 30 days after a chat goes idle |
| Application logs and metrics | Grafana Cloud | IDs and metadata only, no message text, SQL, or tool arguments; about 90 days |
| Files Waldo creates | My Files, in your Organization | Until you delete them |
| Usage records (token counts) | Wherobots billing records | Kept for billing; no chat content |

## Deleting Waldo data

Deleting a user or Organization removes its Waldo chats from the product database. Backups and workflow history age out within 30 days. AI traces, backups included, age out within 90 days of when they were recorded. To have chats or traces removed sooner, contact [support@wherobots.com](mailto:support@wherobots.com). Archiving a chat doesn't delete it.

## Who at Wherobots can see chats

A small group of Waldo engineers reviews AI traces to fix errors and improve answer quality, for example by fixing Waldo's instructions or adding dataset descriptions. This isn't model training. Access to production systems follows the same controls as the rest of Wherobots Cloud: dedicated accounts, SSO, MFA, and logged, regularly reviewed access.

## Compliance

Waldo runs on the same AWS infrastructure, access controls, and change process covered by the Wherobots SOC 2 Type 2 report, and it's in the scope of the Wherobots continuous penetration testing. For security and compliance documents, see the [Wherobots Trust Center](https://trust.wherobots.com) or the [Security Guide](https://docs.wherobots.com/security-guide.md), or contact your account team.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.