XGEN
15-day free trialStart XGEN right in your browser — nothing to installStart free trial

Products · Security & Governance

Tested through an attacker's eyes,
governed the way you declared

XGEN controls and traces AI agent permissions, data access, approvals, and execution. We re-check these controls with penetration testing in an isolated environment to see whether they hold under attack.

Layered Control

Layered guardrails on three axes

XGEN's controls run on three axes: harm review by an external guard model, PII detection and masking, and blocklist detection and masking. Each axis can be switched on or off independently, and detection events are always logged.

Guard Model

Guardrail model

An external guard model reviews user input for harm before it reaches the LLM, and unsafe requests are blocked before any call is made.

PII Masking

Personal data detection and masking

Regex policies detect phone numbers, national ID numbers, emails, account numbers, and similar data, masking them before the text reaches the LLM or gets embedded.

Forbidden Words

Blocklist detection and masking

Regex policies detect and mask the terms your organization has defined. Applied after PII masking, so nothing already hidden gets processed twice.

Guard Model

Harmful requests stop before they reach the LLM

The guard model activates at the workflow agent node and inspects user input before it goes to the LLM. If the request is judged unsafe, the LLM is never called and a block message is returned instead.

Detection categories

ViolenceSexual contentIllegal activityPersonal dataSelf-harm / suicideUnethical conductPolitically sensitive topicsCopyright infringementJailbreak

Strict filtering mode

Switched on, borderline (controversial) results also count as blockable. Switched off, only clearly unsafe requests are blocked.

Security-first operation (fail-open)

Choose what happens when the guard model call fails. In a security-first environment you turn the pass-through option off, so requests stay blocked even during an outage.

Data Protection

Personal data and blocked terms, removed from the source text

PII and blocklist policies run on regex and are managed by policy name, pattern, active state, mask string, and version. During workflow execution, PII masking runs first and blocklist masking second, so an area already hidden isn't processed again.

During workflow agent execution

Masking is applied to the user's input text before it goes to the LLM. Where an internal frame exists (system labels, JSON keys), it applies only to the user-supplied values, preserving the structure.

On document upload

A pre-check shows which items match your policies, and only the policies you select are masked before the actual upload and embedding.

When OCR is used

Text pulled out by OCR gets a supplementary pass of personal-data masking as well.

Pre-detection (detect-only)

Before the user sends any text, the server checks whether personal data or blocked terms are present — without exposing the regex patterns — and shows a warning. It flags the risk at the input stage, with no masking and nothing written to the log.

Audit & Compliance

Every control event, auditable

Events detected or blocked by PII, blocklist, or guard-model policies are written to a unified policy event log. You can filter by source, policy type, user, workflow, collection, and period, and policy creation, edits, and deletions are kept as change history with versions.

Detection source

Workflow agent · execution metadata · Guarder · document upload · OCR supplement

Policy type

PII · blocklist · Guarder

Match details

Matched policy ID and name, detection count, sample source text and masked result

Execution context

Workflow ID · execution ID · interaction ID · node ID

Target

Collection name · document name

User / session

User details · session ID · created at

AI Risk Governance

Managing workflow risk by grade

The AI risk-grading policy is not real-time masking. It is the reference policy for assessing the risk of a workflow or AI service, sorting risk into four levels on four principles and setting the control level accordingly.

Core principles

  • Legality — Compliance with applicable law and regulation
  • Reliability — Accuracy and consistency of results
  • Good faith — Fair and accountable operation
  • Security — Data protection and security controls

Four risk grades

  • Critical — Highest risk level — requires separate controls and an approval process
  • High — High risk level — subject to enhanced monitoring
  • Medium — Moderate risk level — standard controls apply
  • Low — Low risk level — baseline controls apply

On-Premise Architecture

The controls are completed by the on-premise security architecture

Every control policy above assumes air-gapped, on-premise deployment. An external request reaches internal services, AI models, and data only after passing the authentication gateway and the trust boundary, and every model and every piece of data runs on the internal network.

Trust-boundary access

An external request reaches internal services only after the authentication gateway converts its JWT into trusted headers. Clients cannot inject permission headers themselves.

Internal network and air-gap

From LLM serving, embedding, and guard models through to the vector DB and document storage, every AI model and every dataset runs on internal endpoints.

Policy applied consistently across layers

Guard models, PII masking, audit logs, and risk grading are not bound to one layer — they apply consistently across the whole request path.

More on the on-premise security architecture

Attacker's View

Tested through an attacker's eyes

Permission controls are not finished once they are configured. XGEN attempts what an attacker would: reaching another user's data, or internal functions without logging in.

Access Control Test

Permission bypass checks

Automated API tests, configuration scans, LLM probes, and test accounts check whether one account can reach another's data.

Autonomous Pentest

Autonomous penetration testing

ARTEX, an open-source tool in which AI plans its own attack sequence, re-tests authentication and authorization paths.

Isolated Environment

Testing in isolation

A development build runs in a separate environment cut off from external networks, so the shared development server is not affected.

Re-test results

0
Recurrences of prior high-risk items
96.5%
Unauthenticated requests blocked
0
Sensitive data exposures

Results are from a development build tested in early November 2026 and apply only to that build and test scope. Of 1,341 unauthenticated requests, 1,294 were blocked; the other 47 were public status data or responses to malformed requests and nonexistent addresses. Three remaining improvement items are in progress, and validation of the LLM and RAG areas continues.

FAQ

Frequently asked questions

What layers make up XGEN's guardrails?

Three axes: harm review by an external guard model, regex-based PII detection and masking, and blocklist detection and masking. On top of those sits a governance layer — pre-send detection (detect-only), a unified audit log, and the AI risk-grading policy.

When is PII masking applied?

Before a workflow agent calls the LLM, before a document is uploaded and embedded, and again on text extracted by OCR. Detection runs on regex policies, and actual masking happens when the global feature toggle is on.

What happens if the guard model fails?

A fail-open option lets you choose the operating policy. In a security-first environment you turn it off, so requests are still blocked when the guard model is down.

Can control events be used for audit?

Events detected or blocked by PII, blocklist, or guard-model policies are written to a unified policy event log and can be filtered by source, policy type, user, workflow, collection, and period. Policy creation, edits, and deletions are also kept as change history with versions.

How does XGEN test its security?

Automated API tests, configuration scans, LLM probes, and test accounts check whether one account can reach another's data. ARTEX, an open-source tool in which AI plans its own attack sequence, then re-tests authentication and authorization paths. Testing runs on a development build in a separate environment cut off from external networks.

What does the 96.5% block rate mean?

In the early November 2026 re-test, 1,294 of 1,341 requests sent without logging in were blocked. The other 47 were public status data or responses to malformed requests and nonexistent addresses.

How far do the re-test results apply?

Results apply only to the development build and test scope at that time; the security of a production environment can vary with its configuration and operating policies.

Trusted Enterprise AI

Run AI on controls you can verify

Designed for finance and public-sector environments where data sovereignty and regulatory compliance are the starting point — the same control policies apply on-premise. Get in touch whenever you want to review adoption or work through security requirements.