Features How It Works Pricing Whitepaper Get Started Free →
Trust Center · v1.0 · 2026

Zero Trust at the AI Boundary Security & Deployment Guide

The complete reference on data flows, network boundaries, and compliance posture for SecureAIFlow's on-premises and SaaS deployments. Written for security architects, compliance officers, and procurement teams.

Author: BOUAOUDA Achraf · Version: 1.0 · Year: 2026
Zero
Sensitive data retained by SAF
< 500ms
End-to-end detection latency
2 Models
On-premises & SaaS deployment
7+
Frameworks: GDPR · PCI · SOC 2 · ISO · NIS2 · EU AI Act

Abstract

This document describes the security architecture, data flows, and network boundaries of SecureAIFlow (SAF) across its two supported deployment models: on-premises and SaaS. It is intended for security architects, compliance officers, and procurement teams evaluating SAF for regulated enterprise environments.

The document covers: (1) the threat model and scope of detection, (2) the shared security responsibility model, (3) data flow and network boundary specifications for each deployment model, (4) on-premises integration points, (5) regional deployment for SaaS, (6) compliance posture for applicable regulatory frameworks, and (7) a security controls summary.

NOTE

This document does not disclose proprietary detection methodology, model architecture, or internal implementation details. Those aspects are available under non-disclosure agreement upon request.

01

Introduction and Scope

AI-assisted tools — embedded in IDEs, web browsers, and CI/CD pipelines — transmit user prompts to third-party Large Language Model (LLM) providers. These prompts routinely contain source code, configuration files, environment variables, and business data in which credentials, API keys, secrets, and personally identifiable information (PII) may be embedded. This transmission path exists outside the access controls, audit logging, and data classification policies that govern conventional workflows.

SecureAIFlow addresses this gap by intercepting prompts before transmission, identifying sensitive material, and replacing it with deterministic pseudonyms. The LLM provider receives the semantic content of the prompt without receiving the sensitive values. Responses are post-processed to restore original values within the customer's environment.

1.1 Intended Audience

  • Security architects and engineers evaluating SAF for enterprise deployment
  • Compliance officers assessing regulatory alignment (GDPR, PCI DSS, SOC 2, ISO 27001, NIS2, EU AI Act)
  • Procurement and legal teams conducting vendor due diligence
  • CISOs reviewing security controls and the responsibility boundary

1.2 Scope of Detection

SAF detection covers two primary categories of sensitive material:

Category Examples Primary Exposure Context
Credentials and secrets API keys, passwords, access tokens, private keys, connection strings, service tokens All contexts: IDE, browser, SAF UI, automated pipelines
PII (via integration) Customer identifiers, financial references, emails, domain-specific sensitive fields defined by the customer Non-IDE contexts primarily: direct browser interaction with AI interfaces, SAF web UI, support and analyst workflows
NOTE

PII detection is available through optional integration with customer microservices or databases. SAF does not perform PII detection by default without this integration. See Section 4.4 for integration details.

1.3 User Base

SAF is designed for use across all organizational roles that interact with AI systems, not exclusively software developers:

  • Developers and engineers: IDE-integrated workflows, code review, configuration management
  • Data analysts: Direct browser interaction with AI interfaces; may paste data extracts containing customer identifiers or financial fields
  • Support agents: AI-assisted ticket resolution; may include customer PII or account credentials in prompts
  • Operations and DevOps teams: Automated pipelines forwarding configuration or infrastructure data to LLMs

1.4 Interaction Contexts

Sensitive data exposure through AI prompts occurs across the following interaction contexts:

  • IDE-integrated AI assistants: Tools receiving code files or selections as context (VS Code, JetBrains, Cursor, Windsurf, GitHub Copilot)
  • Browser-based AI interfaces: Direct use of ChatGPT, Claude, Gemini, and equivalent by any organizational user
  • SAF web UI: SAF's own interface for multi-LLM interaction, in which users may paste prompts containing sensitive data
  • API-driven pipelines: Automated workflows forwarding application code, configuration, or data to LLMs for analysis or transformation

Traditional secret scanning tools operate at the repository commit layer and do not address any of these real-time transmission contexts. SAF is designed specifically for this gap.

02

Threat Model

2.1 Primary Threat

The primary threat addressed by SAF is the inadvertent transmission of sensitive material embedded in prompts to LLM providers operating outside the organization's security perimeter.

Attribute Description
Threat typeData exfiltration through AI interaction channel
ActorUnintentional (any AI user) — not necessarily malicious
VectorLLM API request containing embedded credential or PII material
TargetAPI keys, passwords, private keys, access tokens, customer PII, financial data
ImpactSensitive data exposure to third-party infrastructure; potential downstream breach or regulatory violation
Existing controls gapSecret scanners operate at git commit layer only; DLP tools cover file/email, not AI prompt layer

2.2 Sensitive Data Categories

Table 1 — Credential Categories in Scope

Category Examples
Cloud provider credentialsAWS access/secret keys, GCP API keys, Azure credentials, DigitalOcean tokens
AI and LLM API keysOpenAI, Anthropic, Hugging Face, and equivalent provider keys
Source control tokensGitHub PATs, GitHub OAuth tokens, GitLab PATs
Payment platform keysStripe live/test keys and equivalent
Communication platform keysTwilio SIDs, Slack bot tokens, SendGrid keys, Mailgun keys
Infrastructure credentialsHashiCorp Vault tokens, Kubernetes service tokens, SSH/RSA private keys
Database credentialsConnection URIs (PostgreSQL, MySQL, MongoDB, Redis), standalone passwords
Session and auth tokensJWTs, internal API tokens, Basic Auth headers, application secrets
PII (via integration)Customer identifiers, emails, financial references — configured per deployment
03

Shared Security Responsibility Model

Security responsibilities are divided between SecureAIFlow and the customer. This division varies by deployment model. The tables below define the boundary for each model.

3.1 On-Premises Deployment

SecureAIFlow Responsible For Customer Responsible For
Detection algorithm accuracy and correctnessVM and infrastructure provisioning and hardening
New version releases and detection rule updatesRules configuration
Integration connector software (VS Code, browser, API endpoint)Network configuration, firewall rules, and perimeter security
Documentation, security guidance, and release notesData residency
Responding to reported detection issuesIntegration configuration (secret manager, database, internal API)
Support
NOTE

SecureAIFlow provides software releases and detection updates. Patching of the operating system, virtual machine infrastructure, and third-party dependencies within the customer environment is the customer's responsibility.

3.2 SaaS Deployment

SecureAIFlow Responsible For Customer Responsible For
Detection algorithm accuracy and correctnessAPI key management and rotation
New version releases and detection rule updatesUser identity and authentication
SaaS infrastructure availability, patching, and securityAcceptable use policy enforcement for all AI users
TLS encryption of all data in transitInterpretation and use of KPI dashboard data
Zero data retention enforcement (technical controls)Integration configuration for PII-aware detection
Regional data residency (customer region assignment)
Anonymized KPI metrics and customer dashboard
NOTE

In SaaS deployment, SecureAIFlow acts as a data processor under GDPR Article 28. A standard Data Processing Agreement is incorporated by reference into the Terms of Service and applies automatically for all customers. A copy is available at secureaiflow.com/legal-dpa.

04

On-Premises Deployment

In on-premises deployment, the SAF detection engine runs exclusively within the customer's infrastructure. SecureAIFlow has no network access to the customer environment and no visibility into prompts, credentials, or audit logs. The customer assumes full infrastructure responsibility as defined in Section 3.1.

4.1 Architecture Overview

Figure 1 — On-Premises Deployment Topology

CUSTOMER SECURITY PERIMETER LLM PROVIDERS INTEGRATIONS · API · DB VAULT AWS SM K8s DOCKER Developer / Analyst AI SAF Engine Local DB · Audit C Claude Gemini AI ChatGPT redacted only

All sensitive data and pseudonym mappings remain inside the customer perimeter. Only redacted prompts cross to LLM providers.

4.2 Data Flow Specification

Table 2 — On-Premises Data Flow

Step From → To Data Content Boundary Status
1Tooling → SAF EngineRaw prompt (code, config, env vars, business data)Within customer perimeter
2SAF Engine ↔ Secret ManagerVault lookup for managed credentials (optional)Within customer perimeter
3SAF Engine ↔ Customer DB / MicroservicePII schema or identifier lookup (optional)Within customer perimeter
4SAF Engine → Local StoragePseudonym map: token ↔ original valueWithin customer perimeter
5SAF Engine → Audit LogDetection event: timestamp, pseudonym ID, detection stageWithin customer perimeter
6SAF Engine → LLM ProviderRedacted prompt: credentials and PII replaced by pseudonymsCrosses boundary — no sensitive content
7LLM Provider → SAF EngineLLM response containing pseudonym referencesCrosses boundary — no sensitive content
8SAF Engine → ToolingRestored response: pseudonyms replaced with original valuesWithin customer perimeter

4.3 Network Boundary Definition

Table 3 — On-Premises Network Boundary

Data Element Crosses Boundary Rationale
Credential values (secrets)NeverPseudonymized before any outbound transmission
PII fields (when integration active)NeverPseudonymized before any outbound transmission
Raw prompt contentNeverProcessed locally; redacted version forwarded
Source codeNeverContained within raw prompt; same boundary rule applies
Pseudonym map and keysNeverCustomer-controlled local storage only
Audit logsNeverWritten to customer-controlled storage
SAF software telemetryNeverNo telemetry transmitted in on-premises configuration
Redacted prompt (pseudonyms only)Yes — to LLM providerNo sensitive content; pseudonyms are opaque tokens
LLM response (pseudonyms)Yes — from LLM providerNo sensitive content; restoration occurs locally

4.4 Integration Points

On-premises deployment supports the following integration points. All integrations are optional and configured by the customer.

Table 4 — On-Premises Integration Points

Integration Supported Systems Purpose
Secret managerHashiCorp Vault, AWS Secrets Manager, Azure Key VaultDeterministic detection of all managed credentials with 100% precision. Any value registered in the vault is detected regardless of format or encoding.
Customer REST APICustomer-defined internal LLM endpointsRoutes redacted prompts to internal or self-hosted LLM endpoints instead of third-party providers. Enables fully air-gapped deployments.
Database / microservicePostgreSQL, MySQL, internal microservicesEnables PII-aware detection by providing SAF with customer-defined sensitive field schemas or identifier patterns. No customer data is copied to SAF — lookup only.
Audit log persistencePostgreSQL, MySQL, or file systemPersists SAF detection events to customer-controlled storage for compliance, forensic analysis, and SIEM integration.
NOTE

All integration traffic remains within the customer perimeter. No integration data is transmitted to SecureAIFlow infrastructure.

4.5 Security Controls — On-Premises

  • Pseudonymization: Credential and PII values replaced with deterministic opaque tokens before any network transmission. Mapping maintained in customer-controlled storage.
  • Audit logging: All detection events logged locally with timestamp, detection stage, and pseudonym reference. Logs contain no original sensitive values.
  • Secret manager integration: When connected, achieves 100% precision on vault-registered credentials regardless of encoding or format.
  • PII integration: Optional lookup-based integration with customer databases or microservices for domain-specific PII detection. No customer data stored by SAF.
  • Air-gapped operation: When integrated with a customer REST API endpoint, all LLM traffic can be contained within the customer perimeter with zero internet egress.
  • Zero telemetry: No usage, diagnostic, or operational data is transmitted to SecureAIFlow infrastructure in on-premises configuration.
  • Latency: End-to-end detection and pseudonymization completes within 500 milliseconds on a standard CPU virtual machine. No GPU required.

4.6 Compliance Posture — On-Premises

Framework Relevant Requirement SAF On-Premises Control
GDPR Art. 32Technical measures for data securityCredentials and PII pseudonymized before leaving the data controller's environment. No personal data transmitted to SAF.
GDPR Art. 44Data transfers to third countriesAll sensitive data processing occurs within the customer's jurisdiction. No credential or PII data leaves the EU boundary.
PCI DSS Req. 3Protection of cardholder dataCredentials never stored by SAF. Pseudonymization is in-memory; mapping is customer-controlled.
PCI DSS Req. 12Audit trailDetection event logs generated locally under exclusive customer control.
SOC 2 (Confidentiality)Protection of confidential informationSAF has zero logical access to customer sensitive material in on-premises deployment.
ISO 27001 A.8Information asset managementCredentials and PII classified and pseudonymized before any external interface. SAF is a software tool, not a data processor.
NIS2 Art. 21Cybersecurity risk measuresPrompt interception prevents exfiltration through AI supply chain. Full audit trail under customer control.
EU AI ActHuman oversight requirementsInterception layer maintains governance over sensitive data flow through AI interactions.
05

SaaS Deployment

In SaaS deployment, prompts are routed through SecureAIFlow-hosted infrastructure for detection and pseudonymization before being forwarded to the LLM provider. The security guarantee is zero data retention: SAF processes prompts in transit but does not persist any prompt content, credential values, PII, or source code fragments. SaaS infrastructure is provisioned within the customer's geographic region.

5.1 Architecture Overview

Figure 2 — SaaS Deployment Topology

CUSTOMER SECURITY PERIMETER LLM PROVIDERS Developer / Analyst AI SAF SaaS · EU/US in-memory only · no storage C Claude Gemini AI ChatGPT HTTPS · TLS 1.2+ redacted only

SAF sits outside the customer perimeter. All processing in-memory · zero prompt content persisted · region-bound.

5.2 Regional Deployment

SecureAIFlow SaaS infrastructure is deployed within the customer's geographic region. Region assignment is determined at account provisioning and is not changed without customer consent.

Customer Region SAF Infrastructure Region Cloud Location Cross-Region Transfer
European Union (France)eu-west-3Paris, FranceNone
United Statesus-centralUnited States CentralNone
Other regionsOn requestDetermined at provisioningNone
NOTE

All prompt processing occurs within the assigned region. No prompt data, credential values, or PII is transferred between regions at any point in the detection pipeline. EU customers' data does not leave the European Union.

5.3 Data Flow Specification

Table 5 — SaaS Data Flow

Step From → To Data Content Persistence at SAF
1User tooling → SAF APIRaw prompt over HTTPS/TLS within assigned regionNot persisted — in-memory only
2SAF API → Detection EnginePrompt for analysisNot persisted — processing only
3Detection Engine → MemoryCredential / pseudonym mappingNot persisted — request-scoped; discarded at end of request
4SAF API → LLM ProviderRedacted prompt: pseudonyms replacing sensitive valuesNot persisted at SAF
5LLM Provider → SAF APILLM response with pseudonym referencesNot persisted — processed in transit
6SAF API → User toolingRestored response with original valuesNot persisted at SAF
7Detection Engine → KPI StoreRequest count, detection rate, latency (no content fields)Persisted — anonymized metrics only

5.4 Network Boundary Definition

Table 6 — SaaS Network Boundary

Data Element Retained by SAF Rationale
Credential valuesNeverPseudonymized in-memory; discarded after processing
Custom fieldsNeverPseudonymized in-memory; discarded after processing
Prompt content (raw)NeverIn-memory only; not written to storage
Source code fragmentsNeverContained in prompts; same in-memory-only rule
Pseudonym mappingNeverDiscarded at end of request; not persisted
LLM responsesNeverProcessed in-memory; not logged or stored
Cross-region data transferNeverAll processing within customer-assigned region
Anonymized KPI metricsYesRequest count, detection rate, latency — no content fields

5.5 Zero Data Retention Controls

  • In-memory processing only: All prompt content is held in request-scoped memory. No write operations to disk, database, or object storage are performed for prompt content, credentials, or PII.
  • Request isolation: Each request is processed in an isolated memory context. Pseudonym mappings are not accessible after the request lifecycle ends.
  • No prompt logging: Application logs do not record prompt content, credential candidates, PII values, or pseudonym mappings. Only operational metadata is logged.
  • KPI metrics are content-free: Retained metrics contain no content fields — aggregate counts, rates, and durations only. No credential type, variable name, PII field, or value is included.
  • TLS in transit: All communication between user tooling and SAF API, and between SAF API and LLM provider, is encrypted using TLS.
  • Regional containment: All processing occurs within the customer's assigned region. No cross-region data transfer occurs at any step.

5.6 Compliance Posture — SaaS

Framework Relevant Requirement SAF SaaS Control
GDPR Art. 28Data processor obligationsSAF acts as processor; DPA required for EU customers. Processing limited to detection only; no secondary use of data.
GDPR Art. 32Security of processingTLS encryption in transit; in-memory-only processing; zero persistent storage of personal data.
GDPR Art. 5(e)Storage limitationPrompt content not retained beyond processing transaction. Retention period for sensitive content: zero.
GDPR Art. 44International transfersEU customers are served exclusively from Paris (eu-west-3) or another selected EU region. No data leaves the European Union.
PCI DSS Req. 3Cardholder data protectionCredential values pseudonymized in-memory; not persisted. SAF retains no payment credential data.
SOC 2 (Confidentiality)Protection of confidential informationZero data retention is auditable: no prompt logs exist to be breached or subpoenaed.
EU AI ActTransparency and data governanceRegional containment and prompt interception support human oversight and governance obligations.
06

Deployment Model Comparison

Table 7 provides a side-by-side comparison of both deployment models across key security, compliance, and operational dimensions.

Table 7 — Deployment Model Comparison

Dimension On-Premises SaaS
Sensitive data retained by SAFNever — not transmitted to SAFNever — in-memory only
Prompt content retained by SAFNever — not transmitted to SAFNever — in-memory only
Data residency100% within customer perimeterSAF-hosted, customer-assigned region
EU data boundary (GDPR Art. 44)Fully controlled by customerEnforced — EU customers on Paris region only
SAF access to sensitive dataZero — no network pathZero — in-memory processing only
Pseudonymization key ownershipCustomerCustomer (per-session; not persisted at SAF)
Audit log ownershipCustomer — local storageCustomer dashboard (KPI metrics only)
GDPR data processor roleNot applicableSAF signs DPA as processor
Secret manager integrationSupported (Vault, AWS SM, Azure KV)
PII detection (via integration)Supported
Air-gapped / internal LLM routingSupported via customer API integrationNot applicable
Infrastructure managed byCustomer IT teamSecureAIFlow
SAF responsible for patchingApplication layer onlyFull infrastructure
Deployment effortMedium — VM provisioning requiredLow — API key configuration only
Latency< 500ms on standard CPU VM< 500ms + network round trip
Recommended forRegulated industries; strict data residency; internal LLM routingTeams requiring rapid deployment with strong regional privacy guarantees
07

Security Controls Summary

Table 8 provides a concise security controls reference for procurement review and vendor questionnaire completion.

Table 8 — Security Controls Summary

Control On-Premises SaaS Notes
Credentials pseudonymized before transmissionYesYesApplies to all contexts and user types
PII pseudonymized before transmissionYes (via integration)Requires DB / microservice integration
Sensitive data persisted by SAFNoNoNever written to any SAF storage
Prompt content persisted by SAFNoNoOn-prem: not transmitted. SaaS: in-memory.
Encryption in transit (TLS)Customer-managedTLS 1.2+SAF enforces TLS for all outbound API calls
Regional data residencyCustomer perimeterCustomer-assigned regionEU: Paris. US: US Central.
Cross-region data transferNeverNeverEnforced by regional infrastructure assignment
SAF access to customer dataNoneNone (processing only)No persistent access in either model
Audit log available to customerYes — full event logYes — KPI metricsOn-prem: full. SaaS: anonymized.
Secret manager integrationYesVault, AWS SM, Azure KV
Air-gapped / internal LLM routingYesNoVia customer API integration
Zero data retention guaranteeN/A — data not transmittedYes — technical controlsSee Section 5.5
GDPR DPA availableNot requiredYes — required for EUContact for DPA
SAF infrastructure patchingCustomer responsibilitySAF responsibilitySee Section 3
CPU-only inference — no GPU requiredYesYes< 500ms end-to-end

Compliance Frameworks Addressed

GDPR
Art. 5, 28, 32, 44
PCI DSS
Req. 3, 12
SOC 2
Confidentiality
ISO 27001
Annex A.8
NIS2
Art. 21
EU AI Act
Human oversight
DORA
Art. 28
08

Notices

This document is provided for informational purposes only. It represents SecureAIFlow's current product architecture and practices as of the publication date, which are subject to change without notice.

This document does not create any warranties, representations, contractual commitments, or assurances from SecureAIFlow. The responsibilities and liabilities of SecureAIFlow to its customers are controlled by the applicable service agreement.

Customers are responsible for making their own independent assessment of this document and for determining its applicability to their regulatory obligations. This document does not constitute legal or compliance advice.

© 2026 SecureAIFlow. All rights reserved. · secureaiflow.com

Want to go deeper?

Download the full PDF, request a DPA, or book a 30-minute security review with our team.