Skip to content

sanitized-audit-report.md

Latest commit

 

History

History
81 lines (64 loc) · 4.25 KB

sanitized-audit-report.md

File metadata and controls

81 lines (64 loc) · 4.25 KB
Raw

Sanitized API Relay Audit Report

This is a sanitized fixture-style example, not a real relay result.

It is built from docs/examples/sanitized-audit-report.fixture.json and follows the draft report artifact schema in docs/report-artifact-schema.md. It is intended to show the shape of a local API Relay Audit report without exposing real relay domains, API keys, wallet material, or private traffic.

Report Artifact

Field Value
Schema version report-artifact-v0.1
Tool version v2.3
Profile full
Target host example.invalid
Overall rating MEDIUM
Generated at 2026-06-02T00:00:00Z
Fixture source Deterministic fixture, not a live relay result

Risk Summary

  • Overall verdict: MEDIUM
  • Prompt injection: synthetic hidden-token delta shown as an anomaly.
  • Prompt extraction: one synthetic extraction result is shown as an anomaly.
  • Context length: inconclusive, not clean.
  • Tool-call rewriting: clean in this fixture.
  • Error leakage: redacted internal path leak shown as an anomaly.
  • SSE anomalies: clean in this fixture.
  • Web3 wallet checks: clean in this fixture.
  • Upstream channel classifier: inconclusive, not clean.

Step Findings

Step Area Status Verdict Severity Summary
1 Infrastructure recon run informational none The target host is redacted to example.invalid for public documentation.
2 Model list enumeration run informational none The relay returned a small synthetic model list in the fixture.
3 Token injection detection run anomaly medium The fixture shows a moderate hidden-token delta so users can see how prompt injection evidence is surfaced.
4 Prompt extraction run anomaly medium One extraction-style probe returned synthetic hidden instruction text.
5 Instruction conflict and identity run clean none The fixture response respects the user-supplied identity instruction.
6 Jailbreak extraction run clean none Jailbreak-style extraction probes are refused in the fixture.
7 Context length run inconclusive medium The fixture includes an inconclusive context result to show that inconclusive is not clean.
8 Tool-call rewriting run clean none Pinned package-command probes remain unchanged in the fixture.
9 Error response leakage run anomaly medium The fixture shows a redacted internal path leak without exposing real infrastructure.
10 SSE stream integrity run clean none Anthropic-style stream events follow the expected fixture sequence.
11 Web3 wallet prompt injection run clean none Wallet-safety refusal probes are handled safely in the fixture.
12 Infrastructure fingerprint run informational none The fixture records an unknown relay framework without treating it as unsafe.
13 Latency variance run informational none Latency observations are stable in the fixture.
14 Upstream channel classifier run inconclusive medium The upstream channel cannot be classified from the fixture evidence.

What This Example Shows

This fixture demonstrates how API Relay Audit separates:

  • clean: the probe did not find an anomaly in this fixture.
  • anomaly: the fixture includes reviewable evidence for a suspicious result.
  • inconclusive: the fixture could not prove clean or anomalous behavior.
  • informational: the step records context without changing the safety verdict.

What This Example Does Not Prove

  • It does not certify any relay as safe.
  • It does not rank or recommend relay providers.
  • It does not represent a live audit of a real relay.
  • It does not replace a local audit against the relay URL you choose.

Redaction Notes

  • No real API keys are included.
  • No real relay domains are included.
  • No private traffic is included.
  • The target host is example.invalid.
  • All findings are deterministic fixture data.

Future Submission Page Compatibility

This example is intentionally schema-first. The fixture JSON can seed future submission-page tests, while the Markdown report gives README and Pages readers a public-safe 30-second view of the report shape.