Guide

How to audit a Claude API relay safely

To audit a Claude-compatible relay safely, run the audit locally, point it only at the relay URL you intend to test, and review each finding as evidence rather than as a safety certificate.

Safe audit checklist

Use a scoped API key with limited budget when possible.
Run the standalone `audit.py` locally so the key does not pass through a separate web checker.
Save the Markdown report and review red, yellow, and inconclusive items.
Do not publish real relay credentials, private domains, wallet addresses tied to private use, or raw traffic captures.

Command

curl -sO https://raw.githubusercontent.com/toby-bridges/api-relay-audit/master/audit.py
python audit.py \
  --key <YOUR_KEY> \
  --url <BASE_URL> \
  --model claude-opus-4-6 \
  --output claude-relay-audit.md

Claude-specific checks

Area	What API Relay Audit checks
Identity	Non-Claude identity leaks, model substitution wording, and stream model identity where available.
Streaming	Anthropic-style SSE event types, usage monotonicity, thinking signature presence, and stream model consistency.
Prompt path	Token injection, prompt extraction, instruction conflict, and jailbreak behavior.
Tool path	Text-echo package command integrity for tool-call rewriting risks.

Interpreting inconclusive results

Inconclusive means the probe could not establish a clean or anomalous result. It may happen because the relay blocks a request, does not support a format, truncates an answer, or returns an ambiguous response. Do not treat inconclusive as clean.

Example report

For the report shape without real relay data, see the sanitized fixture report.