verified_user Security Practices

Security by Design, Not by Checklist

Pharos Platform was architected security-first from day one. Every customer gets a fully isolated environment with dedicated networking, storage, and encryption. No shared infrastructure. No shortcuts.

Data Handling & Privacy

Your data stays yours. Every customer environment is fully isolated with no cross-customer data sharing.

lock

Per-Customer Data Isolation

Each customer gets dedicated Azure storage accounts and a dedicated Key Vault instance. No shared databases, no shared blob containers, no shared secrets. Your data is physically separated from every other customer.

Dedicated Storage Dedicated Key Vault Zero Shared State
public

EU Data Residency

Deploy in EU Azure regions to keep all data within the European Union. Flexible AI model routing supports strict data residency requirements.

enhanced_encryption

Encryption at Rest

All stored data is encrypted using Azure-managed keys in your dedicated Key Vault. Data is encrypted at rest across all storage accounts, databases, and backups.

history

Full Audit Trails

Every operation is logged: who did what, when, and through which channel. Complete traceability for internal reviews, compliance audits, and incident investigation.

block

No Training on Your Data

Customer data is never sent to AI model providers for training purposes. All interactions with AI models are inference-only. Your data stays in your environment.

Infrastructure & Network Security

Defense in depth: multiple layers of protection from identity to data.

fingerprint
Layer 1

Identity & Authentication

check_circle Microsoft Entra ID via your own tenant: SSO, MFA, and conditional access apply
check_circle Signed, short-lived JWT sessions: every request and webhook re-validated, no long-lived credentials
check_circle Zero-trust: all service-to-service calls cryptographically authenticated
check_circle Per-user access control managed in the portal, with roles and per-pipeline scope enforced server-side, so access can never be changed by editing the app on a device
check_circle Governed admin impersonation: support sessions are time-boxed, stripped of privileged rights, always visibly bannered, and audited with both identities on every action
shield
Layer 2

Network Perimeter

check_circle Per-customer Virtual Network with dedicated subnets
check_circle Dedicated Network Security Groups per subnet
check_circle No shared networking between customer environments
database
Layer 3

Data Protection

check_circle Dedicated Azure Key Vault for secrets and encryption keys
check_circle Encryption at rest across all storage services
check_circle Per-customer storage accounts: no shared data stores

AI-Specific Security

Purpose-built safeguards for the unique risks that come with AI agent systems.

AI Safety Pipeline

Every request passes through multiple validation stages before and after AI processing.

shield

Gatekeeper

Incoming requests are screened for spam, phishing, and malicious content before any AI processing begins.

arrow_downward
psychology

AI Processing

AI agents process the request within isolated, controlled environments.

arrow_downward
fact_check

Fact Checker

Before any response is sent, a dedicated agent validates claims against source data to detect hallucinated or fabricated information. Answers arrive with numbered source citations you can verify.

front_hand

Human-in-the-Loop

AI agents cannot execute write operations (sending emails, modifying records, or taking action) without explicit human approval. Approval fits the channel: an approval card in chat, or an explicit spoken confirmation from a verified caller on a voice call. The human stays in control for all consequential actions.

route

Multi-Model Routing

Route AI inference to different providers based on data residency requirements and risk profiles. Provider portability is verified in production, not just promised: agents can switch AI vendors as a configuration change, and a new model or provider can be added the same way, without new platform code. No single-vendor lock-in for AI processing.

block

No Training on Customer Data

All AI model interactions are inference-only. Customer data is never used for model training, fine-tuning, or improvement by any AI provider.

EU AI Act & Compliance

Article 50 transparency has applied since August 2026, and enterprise buyers ask for proof. Pharos Platform builds the required disclosures into platform code at every channel boundary, so compliance is something the platform does, not something your staff has to remember.

Transparency at Every Channel Boundary

Every outbound channel discloses AI involvement automatically. The disclosure lives in platform code, so it cannot be skipped, forgotten, or prompted away. Each disclosure is also logged, so you can show it happened.

forum

Chat, Messaging & SMS

New conversations open with a clear AI notice and every reply carries the channel’s own native AI-generated label. Text messages disclose on first contact, and in the operator dashboard every AI message is badged, with a persistent AI indicator on the conversation itself.

mail

Email

Every AI-authored message carries a localized, human-readable disclosure plus a machine-readable header for automated processing. Forwards disclose the AI-written note without mislabeling the original sender’s content.

call

Voice

Calls open with a spoken AI disclosure in the caller’s language, across every telephony and browser calling path. If the caller talks over it, the platform repeats it, and every call keeps a record that the disclosure was actually delivered.

fingerprint

Marking Anyone Can Verify

Generated documents carry embedded provenance marks that stay with the file when it is downloaded, forwarded, or converted to PDF: open the document properties in any PDF reader and the mark is right there. The full marking specification is published, and a ready-made verification tool is available free of charge on request. Generated markdown carries an embedded, machine-detectable marker, and files in the document store can carry a provenance column for programmatic queries.

verified

No False AI Labels

Only genuinely AI-generated content is marked. Documents an agent receives, redlines, or forwards on a human’s behalf are never labeled AI-generated, so when auditors or customers see the mark, they can trust it. This is a platform-level rule, not an agent’s judgment call.

history

Full Traceability

Every agent interaction is logged: which model produced the response, which tools it called, which other agents it delegated to, and who approved any action that needed sign-off: a complete, queryable audit trail. The model behind each reply is visible right in the conversation, and a reply can be expanded to show the delegation and reasoning behind it, not just the final answer.

gavel

Your Compliance Case, Ready for Procurement

The platform ships the provider-level transparency the EU AI Act requires, keeps its agent catalog outside the Act’s high-risk categories by design, and follows the EU Code of Practice on Transparency of AI-generated Content as its reference framework. As the deployer, your organization keeps its own Article 50 duties, and Pharos Platform gives you the evidence to meet them. Ask for the compliance documentation and we will walk your team through it.

Read the full implementation, channel by channel arrow_forward

Want to discuss security in detail?

We welcome technical security conversations. Request a demo and we’ll walk through the architecture with your security team.