Privacy Utility and Credential Safety Validation

Document Type: Security Validation Report Date: 2026-05-20 Release: v0.0520.2 Runtime: Production AIOrouter Canada-resident routing path Scope: Bidirectional PII usefulness plus technical-secret redaction

Executive Summary

AIOrouter passed the credential-safety boundary test: synthetic API-key-like values were not exposed in the model response, and character-level questions about the API key could not be answered.

The tests also showed that personal-name context can remain useful after privacy processing, while downstream LLM reasoning about surname and middle-name semantics may still vary. This report validates AIOrouter's privacy/security boundary, not perfect model reasoning.

Marketing-Ready Claim

PII stays useful. Secrets stay hidden.

AIOrouter is designed for a practical privacy tradeoff: user context such as names can remain useful to the model workflow, while API keys, bearer tokens, passwords, and private-key material are removed before they can be used, echoed, or inspected.

What Was Tested

Two Kilo Code style prompts were submitted through the production AIOrouter path:

Scenario Prompt Shape Purpose
Chinese travel prompt Chinese personal names plus a synthetic API KEY ... value Verify multilingual name utility and API-key redaction.
English travel prompt Canadian-style English/French names plus a synthetic API KEY ... value Verify English name utility and API-key redaction.

Both prompts asked the model to answer ordinary name questions and secret-character questions, including the third digit of the API key and the fifth-to-eighth characters of the API key.

The public report intentionally omits the full synthetic key-like values. The important evidence is that the live response showed [REDACTED TECHNICAL SECRET] where the key-like value would otherwise have appeared.

Results Matrix

Test PII Utility Technical Secret Safety Result
Chinese prompt Names remained available enough for user-facing answer context, with some Chinese name-boundary overmatch observed. API-key-like value returned as [REDACTED TECHNICAL SECRET]; character-level API key questions could not be answered. Security Pass with model-reasoning caveat
English prompt Full name remained visible; surname and middle-name inference were imperfect in the downstream model answer. API-key-like value returned as [REDACTED TECHNICAL SECRET]; character-level API key questions could not be answered. Security Pass with model-reasoning caveat

Pass Criteria

Criterion Status Evidence
Raw API-key-like value absent from final answer Pass Response used [REDACTED TECHNICAL SECRET].
API key third digit not answerable Pass Model reported it could not see the key value.
API key fifth-to-eighth characters not answerable Pass Model reported it could not see the key value.
Person-name context remains usable Pass with caveat Names appeared or were restored, but model reasoning over surname/middle-name boundaries varied.
Provider echo cannot rehydrate technical secret Pass Secret placeholders remain one-way and cannot become the original key-like value in the final answer.

What This Proves

What This Does Not Claim

Assurance Evidence

Control Public Evidence
Shared secret-safety policy The same credential-safety boundary is applied across API prompts and AI/tool content surfaces.
Provider-bound protection Technical secrets are removed from model-provider-visible context before routing.
Response-side protection Technical-secret echoes return [REDACTED TECHNICAL SECRET] instead of the original value.
Production release v0.0520.2 production validation completed on 2026-05-20.
Detailed implementation evidence Retained internally and available to qualified enterprise reviewers under NDA.

Verification Summary

The release was validated with targeted privacy/security tests, provider-bound wire checks, build verification, API-key leak scanning, and governance validation. Detailed command output and deployment evidence are retained in the internal release record.

Recommended External Wording

Use this wording on public-facing pages:

AIOrouter preserves useful context, such as names in Chinese and English prompts, while preventing credentials from crossing into model-provider context or returning in model responses. In production validation, API-key-like values were reduced to [REDACTED TECHNICAL SECRET], and the model could not answer character-level questions about the secret.