Skip to main content
This page is intended for application owners, security reviewers, and architects evaluating the SDK. It summarises the relevant aspects of how the SDK handles sensitive content and links to the canonical security policy in the public source repository.

​
Data handling principles

Local-first sensitive checks

Pattern-based PII detection runs entirely inside your process. Raw sensitive values are not transmitted to third-party LLMs as part of governance.

Sanitize before judge

When a judge-style policy needs to inspect prompt or response text, it sees the sanitized copy — never the original. Deterministic local checks always run before any LLM-backed check within the same phase, and a Phase-1 block short-circuits Phase 2 entirely on both sides — the judge is never invoked.

Audit reflects what shipped

Audit previews are sampled from the post-sanitization payload, so the dashboard shows exactly what reached the model — not the raw input.

Fail closed on PII

If the local PII engine errors mid-evaluation, the call is treated as if sensitive content was detected. We err on the side of over-blocking.

​
What never leaves your process as part of governance

  • The raw text of payloads matched by local PII rules, before sanitization.
  • API keys you pass to provider SDKs (those go to the provider directly via the provider SDK, not via the EgisAI SDK).
  • Anything covered by your operator-configured local rules.
What does flow to EgisAI:
  • The post-sanitization preview of the prompt and response (what shipped to the model).
  • The verdict, matched rules, latency, and token usage.
  • The agent and user/session metadata you attach via set_context().
You can review what fields a single audit event contains by inspecting any row on the dashboard’s Requests page.

​
Network footprint

The SDK does not make calls to any other third-party services as part of governance.

​
Secrets handling

  • Treat EGISAI_API_KEY and any provider keys as secrets. Use environment variables, a secrets manager, or your platform’s configuration store.
  • The SDK redacts API keys when echoing them in startup logs.
  • set_context(user_id=...) is a free-form string — you control whether it contains an end-user PII identifier or an opaque token.

​
Retention and review

Audit data is retained according to your contract and is reviewable on the dashboard. For long-term archival, ask your account contact about export options.

​
Reporting vulnerabilities

Please do not use public GitHub issues for security-sensitive matters. The canonical report channel is in SECURITY.md in the source repository. We follow a coordinated disclosure process and will work with you on a reasonable timeline.

​
Architecture review summary

A short summary suitable for architecture reviews:
  • Governance evaluates prompts with respect to your organization’s policies before upstream invocation where applicable.
  • Sensitive-content handling is architected so that raw regulated values are not sent to third-party LLMs as part of the policy enforcement workflows described here.
  • Audit metadata is delivered asynchronously so that latency to EgisAI does not sit on the critical path of any provider call.

​
What’s next

How it works

Step-by-step view of the call path.

Releases

Version history and verifying installed wheels.