Dwaar AI settings
Before anyone on your team asks an AI about your findings, your security lead will ask you one question: what leaves the building? Good. This page answers it, and shows you the switch that turns the whole thing off.
Dwaar AI gives members summaries and explanations of what they can already see, writes queries for them on the Query page, and suggests lineage and products for a person to approve. These settings live under Settings › AI. Reading them needs ai.settings:read, and changing them needs ai.settings:update.
What is sent to the model
It is the same for every tier.
- Only the records in scope of the member's page, under that member's access.
- The findings a filter selects, up to the row limit, with the rest sent only as counts. Or one finding with its observations and events.
- For secrets findings: the rule, the path and the masked value.
Never sent: file contents, credentials, member email addresses, or anything outside the console.
Building a query sends less. The model gets the member's question, the GraphQL schema and its limits, plus the query being followed up on. It gets no records, no organization names and no results.
Suggesting lineage and products sends resource details, not findings. The model gets the names, identities, types and provider details of up to 200 resources the member can see, your products with their descriptions and tags, organization names, existing links, and suggestions someone rejected. It also gets Product context facts, unless you turned that category off. It gets no file contents.
Nothing the model answers is stored on Dwaar's side, with two exceptions. A query a member runs is kept with its question in that member's own Query history until they delete it. A lineage or product suggestion is kept with its reason and evidence, because it waits for a person to review it, and anyone who can read those resources can see it. Each run is audited with its surface, tier, rows sent, tokens and outcome, and you can see recent runs at the bottom of the settings page. On Dwaar-managed keys, the records go to the provider behind each tier under Dwaar's account.
Turn it on or off
Members can use Dwaar AI is one switch. Off stops every run at once, while your settings stay in place. On the Query page, members can no longer ask in plain words, but queries they write themselves still run.
Who can use it comes from roles, through ai:use. Administrator, Security admin, AppSec engineer and Developer have it by default. Take it away with a custom role if you want a narrower audience.
The four tiers
Tiers are named by purpose. Dwaar fills them by default and never says which model is behind one.
| Tier | Purpose |
|---|---|
| Fast | Summaries, explanations and drafts. |
| Reasoning | Deeper analysis when a member asks for it. |
| Scanner | Scan-time judgements on skills, models and licenses. |
| Adversarial | Attack generation in Tests Center, which is in preview. |
Members can only ask for Fast and Reasoning.
Defaults
| Setting | What it does |
|---|---|
| Default tier | The tier a summary or explanation uses unless the member asks for another. |
| Offer the Reasoning tier | Members see Analyse deeper, which costs more per run, and Deep on the Query page. Turning it off makes Fast the default. |
| Rows per summary | 50, 200 or 500. The most severe rows are read in full, the rest only as counts. |
Keep the row cap low until someone tells you a summary missed something. A summary of 50 of your worst findings is usually more useful than one of 500.
Keys
There is one mode per tenant, and today that mode is Dwaar-managed.
Your own keys is not available yet. When it arrives, each tier will bind a credential and a model id, with an option to fall back to Dwaar-managed keys when yours fail. Plan on Dwaar-managed for now.
Budget
Spend tracking is not available yet, so a saved budget is not enforced and the budget controls are disabled. Do not treat AI budgets as a spending cap. You can still watch estimated spend under Usage and limits.
Try this next
Open Settings › AI, scroll to the recent runs, and ask for a summary on your findings list. Then come back and find your run. You will see its surface, tier, rows and outcome, which is the answer you can give your security lead when they ask what happened.