Status: draft structure. Owned by Duston Littlefield.
This page exists to be forwarded. A Solutions Consultant evaluating Stormbright should be able to send this single link to their information-security team and have most of the first round of questions answered without a call.
Two kinds of statement appear below and they are not the same:
- Technical descriptions of how Stormbright works. These were written against the running system, not against a roadmap, and they are accurate as at 24 September 2026. Where a control is designed but not yet operating, it says so in those words.
- Commitments by Dustorm PreSales, marked to be set. These are promises about what the company will do, and only Duston Littlefield can make them. Nothing marked that way has been agreed, and an unfilled marker should be read as "no commitment exists yet", not as an oversight.
Third-party commitments are linked, never paraphrased. A vendor's security posture is their statement to make, and a summary of it in our words is worth nothing to your reviewer.
The short version
Stormbright reads the text of sales calls. It runs entirely on Google Cloud in one United States region, us-central1. Models are called through Google Cloud Vertex AI in that same region and nowhere else. Workspace isolation is enforced in the database rather than in application code. There is no recording, no meeting integration and no audio: transcripts arrive because a person pastes text in.
The two facts most reviewers want early, stated without softening:
- Transcripts are stored in full, in plain text, in Stormbright's database, and the optional scrub does not change that. The scrub governs what the model sees.
- Automated retention deletion is not running yet. Data is retained until someone deletes it by hand.
Both are expanded below.
Where transcripts go
A transcript enters Stormbright one way: a person pastes it into the new-call form. There is no upload from a conferencing tool, no bot that joins meetings, no audio processing, and no transcription. Stormbright never touches a recording.
From there:
| Step | Where it happens |
|---|---|
| Stored | Google Cloud SQL for PostgreSQL 17, region us-central1 |
| Scrubbed, if you chose to | In Stormbright's own process. No network call, no model |
| Graded | Google Cloud Vertex AI, region us-central1 |
| Artifacts stored | The same Cloud SQL database |
The transcript is stored before anything else happens to it, as a row in the database. That is a deliberate constraint: the system requires the original to exist so that a scrubbed copy can point at it, which is why "scrub it in the browser and upload only the clean text" is not an option Stormbright offers.
Infrastructure detail
| Control | State |
|---|---|
| Cloud region | us-central1 (Iowa, United States) for every resource |
| Database network | Private IP only. No public IP address |
| Database transport | Encrypted connections only |
| Database authentication | Google Cloud IAM database authentication |
| Encryption at rest | Google-managed AES-256. See Google's default encryption at rest |
| Encryption in transit | See Google's encryption in transit |
| Customer-managed keys | Not enabled. The capability exists in the infrastructure code and is unused |
| Backups | Enabled, point-in-time recovery on, pinned to us-central1 rather than multi-region |
| Object storage | Buckets exist with public access prevention enforced and uniform bucket-level access on. They currently hold nothing: transcripts live in the database, not in a bucket |
Stormbright cannot offer data residency outside the United States today. There is one region.
Who can read your transcripts
Access is enforced by PostgreSQL row-level security. That matters because it is not a check in application code that a new screen or a mistaken query could route around: the rule is attached to the table, and it applies to every statement.
| Who | What they can read |
|---|---|
| The person who uploaded the call | Their own calls, transcripts and artifacts |
| Their manager | The same, for their direct reports and for teams they manage |
| A workspace administrator | Every transcript in the workspace, raw and scrubbed, plus every artifact, plus the raw prompts and model responses |
| Any other workspace | Nothing |
| Dustorm PreSales staff | See below |
The administrator row is the one to raise with your reviewer early. An administrator at your own company can read every transcript every colleague uploads. This is ordinary for a tool an employer buys, but it is a real disclosure and it belongs in the conversation before the first transcript goes in rather than after.
The raw prompt and raw model response are stored for every agent run, and those contain the full transcript text. They are readable by workspace administrators only, not by the person graded and not by their manager.
Dustorm PreSales staff access
Stated precisely, because the precise answer is better than a comfortable one:
- The access-control rules contain no administrative override. There is no rule anywhere that grants Dustorm PreSales staff a route into customer content, and an automated test inspects the database's own catalogue on every build and fails if such a rule appears. This is checked, not asserted.
- Separately, an internal console runs with a credential that bypasses those rules. Today it reads account records, workspace membership and counts. No code path in it opens a transcript. That restriction is held in place by code review, a lint rule that blocks the import, and automated checks, rather than by the database.
So: a person holding production database credentials at Dustorm PreSales could read your transcripts. The controls are real and they are tested, but they are engineering and process controls. Stormbright does not hold your data in a form that makes operator access technically impossible, and no vendor that can debug its own production system does.
to be set: whether Dustorm PreSales commits to a background-check, least-privilege and access-logging policy for staff with production credentials, and what it says
Audit logging
Privileged operations write to an audit log, including refused ones. Workspace administrators can read their own workspace's entries through a view that excludes Dustorm PreSales platform activity. to be set: whether customers can export the audit log, and retention for it
The model provider, and what it commits to
Google Cloud Vertex AI. Project region us-central1. The models in use are Anthropic's Claude family served through Vertex AI as a partner model, with claude-sonnet-5 as the default.
Calls are authenticated with a Google Cloud service account through Application Default Credentials. There is no model API key held by Stormbright for this path.
Your reviewer should read Google's own commitments rather than our description of them:
| Question | Google's published answer |
|---|---|
| What are the contractual terms, including the AI and ML provisions? | Google Cloud Service Specific Terms |
| What are the generative AI terms specifically? | Google Cloud generative AI terms |
| Is our data used to train models? | Vertex AI generative AI data governance |
| What does the data processing addendum say, and what are the deletion and incident terms? | Cloud Data Processing Addendum |
| Which certifications and audit reports exist? | Google Cloud compliance offerings |
| Can we obtain the audit reports themselves, such as SOC 2 and ISO 27001? | Compliance Reports Manager |
| How is Claude served through Google Cloud? | Anthropic Claude on Google Cloud models |
| What is Google's overall data-handling position? | Google Cloud Trust Center |
Where is us-central1? | Google Cloud locations |
Summarised in one sentence, and only because a reviewer will ask for one: Google states that it does not use customer prompts or responses submitted to its generative AI services to train its models. Confirm that at the links above rather than taking it from this page, and treat the Service Specific Terms and the Data Processing Addendum as the ones that matter, because those are contractual and a documentation page is not. If Google's document and this page ever disagree, Google's document is correct and this one is stale.
Google reorganises its documentation regularly and has renamed several of these pages during 2026. The links above are the canonical cloud.google.com addresses, which Google redirects to the current location; if one lands somewhere unexpected, search Google Cloud's own site rather than trusting a restatement here.
to be set: whether Dustorm PreSales will sign a customer's DPA, offer its own, or both
The scrub
Stormbright can replace real names with fixed labels before a transcript is sent to a model. It is worth being exact about its limits, because it is the control most likely to be over-read by someone skimming.
What it is. Deterministic literal text replacement, performed inside Stormbright's own process. It is not a model. It makes no network call. Running it does not send your transcript anywhere.
How it works. You supply a substitution table: each row pairs a term as it appears in the transcript with one of thirteen fixed labels, such as COMPANY, CUSTOMER, PRODUCT 1 or STAKEHOLDER 1. Matching is case-insensitive and longest-term-first, and the labels are always upper case so a human reading the cleaned transcript can see every place a replacement happened.
What it does not do, each of which is a deliberate design choice rather than a gap:
- It does not infer. Only the exact terms you enter are replaced. A plural, a possessive, a misspelling or a transcription variant needs its own row. If the transcript renders "Duston" as "Dustin", you enter "Dustin".
- It does not replace email addresses or phone numbers. It detects them, along with unrecognised proper nouns, and reports them so you can add them to the table. Nothing in that report is altered in the transcript.
- It does not catch account numbers, addresses, amounts, URLs or dates.
- It does not touch the stored original. The scrub writes a second, derived copy. The original transcript stays in the database, unchanged, with the real names in it, for as long as the call exists.
- It is off by default. A new workspace has no substitution table. With no table, with the scrub switched off, or if the scrub is refused because it would have replaced nothing, the transcript is sent to Vertex AI as written.
The table mapping each label back to the real name it replaced is stored alongside the call, and is readable by exactly the people who can read the transcript itself.
The honest summary for a reviewer: the scrub reduces what the model provider sees. It does not reduce what Stormbright stores, and it is not a de-identification control. If your policy is that customer names must never leave your perimeter, the scrub does not achieve that on its own, because the raw text is already in Stormbright's database by the time it runs.
Business context is optional
This is a security position, in Duston Littlefield's words:
Business context has to stay optional; I'm keeping it that way for security reasons. Some users will only want to be graded on their presales delivery, so the agents must work from the transcript alone.
Business context is an optional description of your company, its products and its differentiators that an administrator may add for the agents to read.
Verified against the code: it is optional in the database schema, nothing requires it before a run can start, and when it is absent the prompt receives an explicit "(none provided)" and the run proceeds normally. The agents genuinely work from the transcript alone.
Only a workspace administrator can create or change it. Every member of the workspace can read it. It is typed or pasted text; there is no document upload.
For a reviewer, the practical consequence is that a team can evaluate Stormbright without putting any company-describing material into it at all.
Retention
Read this section carefully. It describes an intention, not an operating control.
The intended defaults are 365 days for transcripts, 730 days for artifacts and grades, and 90 days for raw prompts and model responses. They are configurable per workspace.
No automated deletion is running. Specifically, and verified in the infrastructure code:
- The purge logic is written and works. A bug that made it structurally impossible was fixed on 22 September 2026.
- The scheduled jobs that would call it are created paused in every environment, and the internal endpoint they would call does not exist yet.
- No database-level schedule exists either.
- Nothing has ever been deleted by a retention process.
- No workspace currently has a retention period set in the database, which in this schema means "keep forever".
When the purge does run, it blanks the transcript text and marks the record purged, keeping the row so that artifacts derived from it keep a valid reference.
Until that job is scheduled and has run, the accurate answer to "how long do you keep our transcripts" is "indefinitely, until someone deletes them by hand". This paragraph stays on the page until that changes.
The 365-day figure is also not yet confirmed by Duston Littlefield, and enterprise buyers commonly want 30 or 90. to be set: confirmed default retention period for transcripts
Deletion
There is no self-service deletion workflow and no committed response time.
A user can delete an individual call in the product. That marks it deleted and removes it from every view; it does not erase the transcript text from the database.
There is no account-deletion flow, no workspace-deletion flow, and no built process for a request to erase or export personal data. Such a request is handled manually by an engineer today.
to be set: committed response time for a deletion or access request, once the workflow exists
What Dustorm PreSales has not done yet
Listed rather than omitted, because a reviewer will find them anyway and it is better that they find them here.
- No SOC 2 report. No ISO 27001 certificate. No third-party penetration test. to be set: audit and certification plans, with dates if any
- No customer-managed encryption keys.
- No data residency option outside the United States.
- No published sub-processor page, and no mechanism for advance notice of a change to the sub-processor list. to be set: sub-processor change notification commitment
- No service level agreement. to be set: whether an SLA is offered, and at what tier
- No documented incident response plan or breach notification commitment. to be set: incident response and breach notification commitment, including a notification window
- No bug bounty or coordinated vulnerability disclosure policy. to be set: security contact and disclosure policy
- No formal business continuity or disaster recovery plan, beyond database backups with point-in-time recovery pinned to the region.
- No organisation-level cloud policy constraint pinning resource locations. The single-region commitment is held by configuration and review, not by a guardrail that would refuse a resource created elsewhere.
Questions this page does not answer
Send them to to be set: security@ address and they will be answered directly rather than guessed at here.
If your reviewer needs something in this page turned into a contractual commitment, that is a conversation with Duston Littlefield, not a change to this document.