The security module thatprotects data before the model.
PASBAN Gateway is the security enforcement module of the PASBAN Enterprise AI Security and Governance Platform. It runs every request through detection, policy, routing and logging, and controls sensitive data and agent risks.
Turkish NER & latency — how we measure, how you reproduce.
F1 (entity-level) + p50/p95/p99 latency. Dataset, labeling guide & scripts are published — claims verified in-table.
| Entity type | N | Precision | Recall | F1 |
|---|---|---|---|---|
| Person name | — | — | — | — |
| National ID | — | — | — | — |
| IBAN | — | — | — | — |
| Phone | — | — | — | — |
| Address | — | — | — | — |
| Sensitive (health) | — | — | — | — |
| Overall (micro) | — | — | — | — |
| Prompt | p50 | p95 | p99 |
|---|---|---|---|
| Short | — | — | — |
| Medium | — | — | — |
| Long | — | — | — |
Dataset version + hash, inter-annotator κ, fixed seeds, bootstrap 95% CI. Separate pilot measurement on your data recommended.
Three layers, one decision input
Entity recognition (NER)
Person, address, health — Turkish morphology-aware model
Pattern & checksum
TC ID, IBAN, phone, tax ID — regex + verification
DLP rules
Source code, customer rows, secrets, API keys — regex + optional AI classifier
Turkish suffixes are treated as part of the entity. For sensitive data, policy may choose block instead of mask. Extensible to 30+ types.
If undecidable, it blocks. No unchecked data flow.
Each row is tested, results will be shared in the security package. Target HA and SLA are scoped in the pilot.
| Case | Behavior | User sees | Log |
|---|---|---|---|
| Entity recognition unavailable | Request blocked | Temporary error, retry | Incident and cause logged |
| Policy engine undecided | Request blocked | Policy error | Incident logged |
| Below confidence, sensitive data | Block by default for sensitive types, mask for low sensitivity + threshold logged | Reasoned notice | Decision and threshold logged |
| Response filter unavailable | Response withheld | Temporary error | Incident logged |
| Audit write fails | Request blocked (503), encrypted local buffer, SIEM alert, no bypass | Temporary error, retry | Buffered then flushed, failure itself logged |
| Model provider unreachable | Fallback only with same residency + equivalent assurance + explicit policy allow, otherwise hold-and-block | Short notice | Routing decision logged |
| Quota exceeded | Rejected (HTTP 429 quota_exceeded + Retry-After) | Quota message | Record kept |
Blocking doesn’t stop use — it hides it.
When enterprise access is closed, employees keep using personal accounts. That’s shadow AI — you lose visibility into which data goes to which model.
PII leaks into chat
Rows with names, IDs and IBANs are pasted into a model to summarize.
Code & secrets leak out
A developer pastes proprietary code with embedded API keys to debug.
No audit trail
No record of who accessed what, with which model, and when.
Security and governance in one architecture.
AI usage grows from chats to agents. PASBAN Gateway governs that growth with three functions, centrally.
Set policy centrally
Which model, which data, which unit is open, restricted or blocked? Policies, RBAC and DLP in one place.
Inventory & budgets
Budgets, quotas and model routing are managed in one place today. The registry for approved models, prompts, MCP connectors and agents joins the same control plane in the next release.
Prove what happened
Who used what, with which model, at what cost? Usage, cost, anomaly and audit trail in one pane, streams to SIEM.
The core control plane is available today. Model, prompt, MCP and agent registry, plus agent governance, join the same architecture in the pilot — no rip-and-replace.
Every request passes four gates before the model.
No request passing through PASBAN Gateway bypasses policy. If undecidable, default is block (fail-closed).
Entity recognition
Turkish-tailored NER finds PII, DLP rules find secrets & code. Prompt injection flagged.
Decide
Unit- and role-aware rules: mask, block, allow or route to on-prem model.
Model selection
Cleaned request goes to public LLM or on-prem model. Quota & budget checked in-line.
Audit trail
Request and response logged with who/when/decision and timestamps. Known secret patterns masked best-effort before write.
Rules are defined per unit and role, one row per rule.
One row per request: policy decision, routed model and outcome logged together.
Control for security, productivity for employees.
PII detection
TC ID, IBAN, phone, address, tax ID, plate, name variants and sensitive PII. Handles Turkish suffixes & context. 30+ types extensible.
Leak prevention
Source code, customer DB rows, trade secrets. Custom regex. Prompt injection & jailbreak detection, output filtering.
Policy engine & audit
Unit- and role-based rules. Timestamped logs are forwarded to SIEM systems (CEF/syslog). OpenTelemetry separate for observability.
Cost, quota & routing
Token rate limiting, per-user/project/model budgets. Stop on quota breach, chargeback/showback, multi-provider routing & fallback.
Identity & access
Directory sync via LDAP/AD sync separately, unit and role RBAC, API key issuance, rotation and revocation.
Secure chat for employees
Chat UI, file upload and analysis, Word/Excel/PDF generation, web search. SharePoint and Jira connectors are optional and disabled by default. Enforcement stays invisible.
Control plane stays with you. Model choice is your policy.
Detection, policy and logging run in your environment. You decide which requests go where.
The list is illustrative — other providers and self-hosted models can be added to routing.
The audit trail never leaves your boundary.
On-prem and VPC logs stay inside your boundary. Timestamped, known secrets masked best-effort, RBAC-gated and self-logged. Logs integration with SIEM systems. OpenTelemetry separate for observability.
- Timestamped · who / when / unit / model / decision
- Known secret patterns masked best-effort before write
- RBAC on access — access itself is logged
Every team has a different requirement.
These four requirements are what teams need before AI can be used safely. PASBAN Gateway meets all of them in one control plane.
Know which model runs on which data
All model access is visible in one pane, unauthorized use and risky data flows surface immediately.
Prove it at audit time
Every request is logged with user, time and decision, personal data is processed before it reaches the model.
Cap spend before it happens
Per-unit budgets and quotas, requests stop automatically on breach and spend is charged back to units.
Keep working with code and customer data
Prompts containing source code or customer records are masked, blocked or routed on-prem by policy.
Supports compliance, but does not replace it.
Helps you meet KVKK, GDPR and EU AI Act with technical controls and evidence. Independent legal assessment is in preparation, it will be shared here once available.
.svg.png){kind=logo}
- ISO/IEC 27001Information Security Management
- ISO/IEC 42001AI Management System
Starts with PASBAN Gateway, expands to an enterprise AI control plane.
Same core architecture. No new module when agents arrive — the control plane expands.
PASBAN Gateway & workspace
PII & DLP, policy engine, audit trail, SSO & RBAC, cost control, multi-provider routing.
Model, prompt, MCP & agent registries
Approved model & MCP inventory, agent authority bounds, human-in-the-loop, MCP Gateway & agent security.
CISOs ask first.
How is masking reversed?+
What if PASBAN Gateway is down?+
How is Turkish detection measured?+
How is it different from Purview or other hyperscaler controls?+
Where are logs stored?+
Which models are supported?+
Which certifications?+
Let’s plan your enterprise AI stack.
We demo with your scenarios: data policies, deployment, and how the platform expands for you.
Request sent!
We'll be in touch shortly.
One more step
Your email app is opened with your request pre-filled. Your request is only sent once you send that email.
Didn't open? Email us directly:contact@pasban.ai