Protect sensitive data in internal AI copilots

Internal copilots are productive precisely because employees paste real context into them: customer emails, account details, internal tickets, sometimes credentials and tokens. Every one of those prompts is then forwarded to a third-party LLM. Most teams have no consistent answer to the question "what did we send to the model today?"

• Sensitive identifiers in prompts

  Names, emails and account references leak into provider context.

• Pasted secrets

  API keys and tokens land in chat boxes and get forwarded along.

• Fragmented policy

  Each copilot ends up with its own ad-hoc data-handling story.

• Hard to audit

  Without a central hop, you can't tell which prompts touched which model.

• Single egress hop

  All internal AI traffic flows through one auditable endpoint.

• Mask before the model

  Supported sensitive values are replaced with placeholders before egress.

• BYOK

  Provider usage runs on your own keys and shows up on your provider account.

• Zero retention

  Raw prompts and responses are never persisted by Privian.

Employee tool  →  Privian gateway  →  LLM provider
                    │  mask
                    │
                (placeholders only)
                    │
                    ▼
                rehydrate  ←  provider response
                    │
                    ▼
                Employee UI

curl https://api.privian.io/v1/gateway \
  -H "Authorization: Bearer $PRIVIAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4.5",
    "prompt": "Summarize this email thread from jane@acme.com about onboarding.",
    "stream": false
  }'

Same small JSON contract for every internal copilot, regardless of which provider it ultimately hits.

• Internal GPT assistant

  Company-wide chat assistant routed through one masking hop.

• Ops copilots

  Runbooks and incident assistants that touch sensitive system data.

• Support copilots

  Internal helpers for support teams handling customer context.

• Engineering copilots

  Coding assistants that occasionally see secrets in pasted snippets.

• No raw prompt or response storage
• Provider credentials encrypted at rest (AES-GCM)
• Gateway API keys stored as SHA-256 hashes
• Provider-agnostic — switch models without changing posture
• In-memory entity mapping, discarded per request
• Structural observability only (counts, never content)

1. Step 1

   Add provider keys

   Register your OpenAI / Anthropic / Google keys in the dashboard.

2. Step 2

   Issue scoped gateway keys

   One key per copilot makes auditing and rotation simpler.

3. Step 3

   Repoint copilots

   Change the base URL to https://api.privian.io. No prompt rewriting.

4. Step 4

   Restrict models

   Use the dashboard to limit which models a given key can call.

5. Step 5

   Verify masking

   Send a known sample through and confirm placeholders in the provider view.

6. Step 6

   Observe and iterate

   Track request volume and entity counts. Tune as new copilots come online.

• Not a replacement for enterprise security programs
• No prompt injection or jailbreak defense
• No OpenAI SDK drop-in — Privian uses a small JSON contract
• No native provider token streaming yet
• No HIPAA / SOC 2 / PCI compliance claims today
• Detection covers the supported entity set only

Can Privian secure internal copilots?

Yes, for the data Privian can see. Every prompt is scanned for the supported entity set and matched values are masked before egress. Internal copilots inherit one consistent privacy layer instead of each tool implementing its own.

Does Privian prevent prompt injection?

No. Privian focuses on prompt privacy — masking sensitive values before they reach the provider — not on detecting or blocking adversarial prompts. Treat injection defense as a separate concern.

How is employee data handled?

Raw prompts and responses are never persisted. The entity mapping that lets us rehydrate the response is held in memory for the lifetime of a single request and discarded afterwards.

What sensitive values are supported?

Personal identifiers (names, emails, phone numbers, IPs), financial and national IDs (cards, IBANs, US SSN, Canadian SIN), and developer secrets (JWTs, OpenAI keys, GitHub tokens, AWS keys, generic API keys, env-style secrets).

Does Privian replace enterprise security?

No. Privian reduces prompt-level exposure of supported sensitive values. Network controls, identity, DLP and audit programs remain your responsibility.

Where are provider credentials stored?

Provider keys are encrypted at rest (AES-GCM) and only decrypted in-memory at request time to call the upstream model. Gateway keys are stored as SHA-256 hashes.

• Product: LLM Gateway
• Product: PII Masking
• Product: Prompt Security
• Guide: Protect internal copilots
• Concepts: Zero retention
• Concepts: BYOK
• Glossary: LLM security
• Category: AI Security Layer