Trust & Security

Security and data protection
built in, not bolted on.

Aegis is a governance layer, so protecting data is not a feature we added — it is the reason the product exists. This page sets out, plainly, how we handle your data today, and where we are honest about what is still on the roadmap.

The six pillars

What protects your data, end to end

Each control runs inline on every request — before a model is ever called, and after every answer is returned.

PII redacted before any model sees it

Deterministic regex and rule packs are the authoritative detector; your own custom tenant rules layer on top. An optional multilingual transformer (off by default) augments them by catching names, organisations and places the patterns miss. Detected values are replaced with placeholder tokens; only the placeholders are sent to the model.

The optional NER model is trained on ten languages including Arabic. Hindi is not one of them; Hindi PII is covered by deterministic rule packs rather than the model.

Encrypted at rest, with per-tenant keys

Message content is encrypted with AES-256-GCM, enabled in production today. Each tenant has its own Data Encryption Key, wrapped by a master Key Encryption Key that is held only in gitignored configuration, never committed to the repository and never stored alongside the data it protects.

Key rotation is not yet automated (one active key per tenant), and rows written before encryption was enabled may remain plaintext until an optional backfill runs.

We never train on your data — opt-in only

Your prompts and outputs are not used to train the shared CLAB model unless you explicitly opt in, and that setting is off by default. If you wish, you can train a private adapter on your own data alone, isolated to your tenant.

Full audit trail & compliance reporting

An append-only audit event is written per request, best-effort: a failed write is logged and never blocks the request. The audit log records the counts and types of PII detected - never the values themselves. Export the trail as CSV or JSON at any time, generate a monthly compliance PDF, and run DSAR export and delete on demand.

Your own keys & your own model

Bring your own upstream provider API key — encrypted at rest with the same envelope as your data. Or run inference entirely on the local CLAB-Aegis model, where nothing leaves the box. Frontier providers, open models and sovereign local inference, behind one gate.

Topic firewall & supervisor approvals

Allow or block topics per tenant, route sensitive requests through supervisor approval workflows, and enforce three-tier quotas across tenant, user and department. A prompt-injection guard and a hallucinated-identifier guard run alongside redaction.

Where your data goes

The journey of a single prompt

Five steps from what a person types to what comes back — with the original sensitive values never leaving your control.

1

Detect

The prompt is scanned the moment it arrives, before any model is contacted. Deterministic rule packs and your custom rules are authoritative; an optional multilingual NER model augments them when enabled.

PiiDetectionStage
2

Tokenise & vault

Each detected value is replaced with a <<PLACEHOLDER>> token. The real value is stored tokenised in an encrypted vault, gated to the owning session.

tokenised vault
3

Send placeholders only

Only the placeholder-bearing text reaches the model — your key, an open model, or the local sovereign model. The original PII is never transmitted.

placeholders → model
4

Re-inject

The answer comes back with placeholders intact. Aegis re-injects the real values from the vault, so the person sees a complete, accurate response.

PiiReinjectionService
5

Audit & encrypt

An append-only audit event is written per request (best-effort; failures are logged, never block the request) - counts and types of PII only. Stored message content is encrypted at rest with your tenant key.

counts & types only

In short: detect → replace with tokens → only tokens go to the model → real values re-injected into the answer. The audit log holds PII counts and types — never the values.

Compliance posture

Designed to support your obligations

We are precise about this on purpose. Aegis provides controls and tooling that help you meet regulatory obligations — it does not, and we will not claim it does, grant you a certification.

Designed to support GDPR, DPDP & PDPL-style regimes

The architecture is built to help you meet data-protection obligations under regimes such as the EU GDPR, India's DPDP Act and GCC PDPL frameworks.

  • Data-minimisation by design — models see placeholders, not personal data
  • Purpose limitation via topic firewall and policy presets
  • A DataRegion attribute recorded per tenant

Audit & data-subject tooling

The practical tools an auditor or a data subject will ask for, available without a support ticket.

  • Append-only audit log; export as CSV or JSON
  • Monthly compliance PDF — detections by type (counts only), outcomes, approvals, alerts
  • DSAR export and delete; public status page at /status.html
What we will not say. Aegis is not certified or attested under SOC 2, ISO 27001, HIPAA, GDPR, DPDP or any other regime, and no third-party security audit or penetration test has been performed. We describe the platform only as designed to support these obligations and as something that helps you meet them. Anyone who tells you otherwise is mistaken.

Current limitations

We would rather you read this from us than discover it later. For serious buyers, candour is part of the security posture.

  • Single-region SaaS today, hosted on the founder's infrastructure — no multi-region or high-availability failover yet
  • No third-party security audit or penetration test has been carried out — it is on the roadmap
  • Encryption key rotation is not yet automated; one active key per tenant
  • Rows written before encryption was enabled may remain plaintext until an optional backfill is run
  • The sovereign on-prem appliance and dedicated-database-per-tenant isolation are designed but not yet built
  • Hindi is not a trained language for the current NER model; Hindi PII is covered by deterministic rule packs rather than the model
Go deeper

Documents & tools

Everything you need to take Aegis through a procurement or security review.

Security whitepaper

The full architecture: the redaction pipeline, the encryption envelope, the data-sovereignty model and the threat model behind them.

Read the security whitepaper

System status

Live availability and incident history for the Aegis platform, with no login required.

View system status

Admin console

Sign in to configure policies, manage keys and quotas, and export your audit trail and compliance reports.

Sign in