Data shared without noticing
A contract, customer spreadsheet, medical report or source file pasted into a chat window leaves your control. Most of the time you cannot see what went where.
The door you open to AI opens both ways.
Derbent AI is the security layer that lets you use AI without exposing your data to external AI models. For individuals, teams and companies.
What is Derbent AI?
Derbent AI is a security layer that sits between your documents, code and accounts and external AI models and agents such as ChatGPT, Claude and Gemini: a two-way AI gateway. Only the context a task requires goes out; only verified actions come in. Whether you work alone, in a team or across a company, you keep using AI while your data never leaves uncontrolled.
Data leakage
Derbent isolates the minimum context a task needs. Instead of the whole document, only narrow, relevant questions reach the model; your documents stay with you.
Agent permissions
Actions by AI agents, such as reading files, changing code or deploying to production, are checked against the scope of the task. Risky actions wait for human approval.
What it knows and what it can access
are not the same thing.
Inside · the user and their documents
research.pdf
experiments.xlsx
Advisor meetingresearch-v2.pdfEdited · localresearch.pdfChapter 3 · 42 pagesexperiments.xlsxUnpublished dataAdvisor meetingMeeting notes
Where should we start?
research.pdfexperiments.xlsxAdvisor meetingCan you edit my academic paper?
Derbent OnelocalWorking
- Read the full context locally 4 documents · never left
- Split the work into four sub-questions
- Released only the questions
- Verified the returned answers
- Merged the result locally
research-v2.pdfEdited · local- References corrected to APA 7
- Equation explanation simplified
- Figure 3 caption clarified
research.pdfChapter 3 · 42 pagesexperiments.xlsxUnpublished dataAdvisor meetingMeeting notesAsk anything
Derbent
Outside · models and agents
ChatGPT
Claude
Gemini
GrokREAD EXPORT SUBMIT
What external models see
“Can you rewrite this paragraph in a more academic tone?”
Never saw the rest of the thesis.↩ Suggested consistent terms and passive voice.
Claude“Is this explanation of the equation clear enough?”
Never saw the experiment data.↩ Define ζ and ω₀ at first use.
Gemini“Does the reference list follow APA 7?”
Doesn’t know the author or institution.↩ DOI format missing in three sources.
Grok“How can the figure caption be clearer?”
Never saw the unpublished figure.↩ Name the axes and the damping ratio.
What the agent requested
Read the draft notes
Advisor meeting
Checking scope and policy…✓ Allowed In scope, read-only.
Export the unpublished data
experiments.xlsx
Checking scope and policy…× Denied This data cannot leave the boundary.
Submit the final version to the journal
research-v2.pdf
Checking scope and policy…Ⅱ Approval required Submission awaits human approval.
The AI never sees your files, only narrowly scoped questions.
Your most valuable context
is the context that must not leave.
research.pdf07 / 18Research notes · draft
The geometry
of stability
of stability
Convergence bounds in damped systems
ẍ + 2ζω₀ẋ + ω₀²x = 0
Figure 3. Phase trajectories converging to equilibrium.
Unpublished research
session.ts
async function capture(id) {
const s = await tx.lock(id);
if (s.paid) return s;
return ledger.capture({
key: s.idempotencyKey,
amount: s.total
});
}
Teams#paymentsDA
Deniz AkınConcurrent payments failing at two checkouts.
SE
Selin ErdemLet’s hold the deploy. Checking the logs.
customers.xlsx| A | B | C |
|---|---|---|
| Ilgın Logistics | ₺842,500 | 30 days |
| Maviova Foods | ₺316,800 | 45 days |
Atlas — next release
OctoberNew payment infrastructure
NovemberClosed customer pilot
This data
must not leave.
Derbent
Race condition analysis
ClaudeKey collision
GeminiLatency pattern
GrokAlternative fix
Documents stay inside. Only scoped questions pass through.
Not just what it sees.
What it can do.
Ask ClaudeInvestigate the payment bug.
Derbent
session.tssrc / auth
reconcile.pyservices / payments
payments-apiproduction
DENIED
Outside the task of investigating the payment bug.
Context stays inside.
Intelligence comes from outside.
Ask the local agent
Find the cause of the payment bug and prepare a safe fix.
customer.xlsx
#payments
session.ts
incident-report.pdf
Cause of the double write found.A fix was drafted locally.
Derbent
If two concurrent transactions read the same record, in what order can a double write occur?
Doesn’t know the company.↩ Make the read and write atomic.
ClaudeUnder what conditions does this idempotency key collide?
Doesn’t know the customer.↩ One persistent key per transaction.
GeminiIs this latency distribution consistent with lock contention?
Doesn’t know the repository.↩ Measure lock wait time separately.
GrokHow can a second write be prevented without a distributed lock?
Doesn’t know the original task.↩ A unique constraint and an atomic transition.
One boundary. Two directions.
Company data and systems
strategy-q4.pptx
customers.xlsx
Customer email
incident-4471.pdf
Derbent
Models / Agents
ClaudeGeminiGrokCompany → Derbent → ModelsMinimum necessary context
Company ← Derbent ← AgentsVerified action
The controlled boundary between your company’s
private environment and AI.
Data protection and data minimization
Data protection laws such as the GDPR and Turkey’s KVKK (Law No. 6698) require personal data to be relevant, limited and proportionate to the purpose it is processed for, and they attach separate conditions to transfers abroad. Most external AI services run in other countries. Derbent’s “minimum necessary context” principle limits the data sent to a model to what the task requires: for individuals, that means less personal data leaving; for organizations, it works in the same direction as the data minimization principle. Derbent does not make you compliant on its own; compliance has to be assessed together with the decisions of the data controller.
Frequently asked questions
- What is an AI gateway?
- An AI gateway is the layer placed between users and applications and AI models that inspects the traffic between them. Derbent builds this layer in both directions: it limits the context going out and checks the agent actions coming in.
- Is Derbent only for companies?
- No. Derbent is being designed for individuals, teams and companies: anyone who wants to use AI without losing control of their data.
- Does Derbent ban the use of AI?
- No. The goal is not to block AI but to let you benefit from it while keeping your data under control.
- Is Derbent available?
- Not yet. The platform is not publicly available; you can follow progress on our LinkedIn page and send questions to [email protected].
Guides: What is an AI gateway? · What is shadow AI?