toby-bridges/api-relay-audit

API Relay Audit

Local security audit for AI API relays and LLM proxies.

DSH Plugin

Your Agent Is Mine: what you can test locally

*Your Agent Is Mine* documents malicious API relays injecting payloads and exfiltrating credentials. describes fraudulent Claude resellers swapping models and harvesting credentials through their client tooling. In his September 11 disclosure, co-author Chaofan Shou reports buying router data containing users' credentials.

API Relay Audit is an independent, local security audit tool for AI API relays and LLM proxies, informed by the paper. The current release checks observable relay behavior and generates a Markdown report covering:

• Prompt and context signals: hidden prompt injection, instruction override,

and context truncation.

• Response integrity: changes to pinned package-command text, error-response

leakage, and SSE stream anomalies.

• Reviewable findings: per-step evidence and LOW / MEDIUM / HIGH summaries;

inconclusive probes remain visible.

See the output: example report (synthetic fixture) · Try it: run a local audit · Coverage and limits

The standalone script uses Python's standard library plus curl. Your API key is sent only to the relay URL you choose.

Quick Start

AUDIT_SCRIPT_REF=v2.5.0
curl -fsSL "https://raw.githubusercontent.com/toby-bridges/api-relay-audit/${AUDIT_SCRIPT_REF}/audit.py" -o audit.py

python audit.py --key <YOUR_KEY> --url <BASE_URL> --output report.md

# Web3 / wallet users
python audit.py --key <YOUR_KEY> --url <BASE_URL> --profile web3 --output report.md

See a public-safe fixture report: sanitized audit report. Pinned v2.5.0 downloads require a published tag; a draft release alone does not guarantee availability. Use master as AUDIT_SCRIPT_REF only when intentionally testing unreleased changes.

If API Relay Audit helps you evaluate a relay before sending real traffic, star the repository to follow new detector coverage and release-tested updates.

JSON Reports

v2.5.0 supports --format json for local automation:

python audit.py --key-env API_RELAY_AUDIT_KEY --url <BASE_URL> --format json --output report.json

Without --output, stdout contains one JSON document and progress goes to stderr. JSON preserves the existing final risk rating and full Markdown evidence; it does not sanitize the report for public sharing. See the JSON schema and coverage boundaries.

When to Use It

• You use a third-party AI API relay, mirror, gateway, or LLM proxy.
• You want to check whether a Claude-compatible or OpenAI-compatible proxy injects prompts, swaps models, truncates context, or rewrites tool output.
• You are testing relay behavior before production traffic, coding-agent automation, package-install suggestions, or wallet-related actions.
• You need a local, repeatable audit report instead of a web tool that asks for your API key.

What It Does Not Claim

• It does not certify that a relay is safe.
• It does not replace manual security review or operational monitoring.
• It does not treat inconclusive as clean; blocked probes and ambiguous responses stay visible in the report.

Query Family Boundaries

Query family · User intent · Profile / steps · Evidence boundary API relay audit · Audit a third-party relay, mirror, gateway, LLM proxy, or resale API before trusting traffic. · general by default; full for every probe · Produces a local report, not a safety certificate. Prompt injection audit · Detect hidden prompt injection, prompt leakage, instruction override, and extraction behavior. · general; Steps 3-6 · Records prompt evidence without publishing private prompts or secrets. Model substitution signals · Collect model identity, stream, latency, and upstream channel signals. · general; Steps 5, 10, 13, 14 · Self-ID, latency, and channel fingerprints are signals, not standalone proof of provider substitution. Web3 relay audit · Check wallet-sensitive relay behavior before agent workflows touch signing or transactions. · web3 or full; Step 11 · Profile-gated; general relay audits do not imply wallet safety.

The canonical contract lives in docs/query-families.md. README headings, Pages cards, issue templates, and skill descriptions should preserve these boundaries instead of flattening them into one slogan.

Coverage

API Relay Audit checks whether a relay modifies the request or response path between you and the model:

• Prompt safety: token injection, prompt extraction, instruction override, jailbreak resistance
• Relay integrity: context truncation, tool-call substitution, error leakage, stream integrity
• Model identity: non-Claude identity leaks, model substitution signals, Claude/OpenAI-compatible relay behavior
• Web3 wallet safety: transfer guidance, signed-transaction refusal, private-key refusal

Audit LLM Proxies Locally

The project has two distribution modes:

• audit.py: zero-dependency standalone script for quick local audits
• api_relay_audit/ plus scripts/: modular development version with tests

Runtime profiles:

• general: default AI API relay and LLM proxy checks
• web3: wallet-safety probes for Web3 agent flows
• full: general plus Web3 checks

DeepSeek Harness DSH Plugin

The repository is also an installable dsh-api-relay-audit bundle for DeepSeek Harness Web and community TUI surfaces that use the official @deepseek-ai/dsh-commands registry. Pin an immutable commit or release tag:

DSH_PLUGIN_REF=v2.5.0
dsh plugin --profile web add "github:toby-bridges/api-relay-audit#${DSH_PLUGIN_REF}"

# dsh-cc-tui and other compatible profile-based clients
dsh plugin --profile cc-tui add "github:toby-bridges/api-relay-audit#${DSH_PLUGIN_REF}"

The command reuses the current DSH provider's baseURL, model, and credential reference. The credential stays in DSH Credentials and is delivered to the local audit process through an environment variable, keeping it out of the plugin command input, the audit process arguments, and the session log:

/relay-audit
/relay-audit --connectivity
/relay-audit --profile web3 --fast-context
/relay-audit --url <URL> --model <claude-model> --credential-ref <DSH_CREDENTIAL_REF>

No arguments preserves the existing full-audit default and may consume metered tokens. Use --connectivity for a lower-cost check. This distribution does not add a new model baseline: the selected route must identify as Claude, although the relay API itself may be Anthropic-compatible or OpenAI-compatible. The older v2.4.0 release can put the key in curl process arguments during error probes and streaming fallback. v2.5.0 passes those headers through curl config stdin instead. Treat the local process list as sensitive if you still run v2.4.0. Independent wrappers without DSH profiles and the DSH command registry are not compatible with this bundle. See agent distribution notes. The exact v2.4.0 installation, runtime, and secret-scan results are recorded in the DSH distribution verification. For the updated host matrix and verification limits, see the v2.5.0 release notes.

Retained Agent Skill Files

The repository retains its existing OpenClaw and Hermes skill files for direct users and downstream compatibility. They are not current registry distribution targets; active distribution and release verification focus on the DeepSeek Harness plugin.

• OpenClaw Skill: run a local AI API relay audit before an OpenClaw agent

depends on a third-party relay, proxy API, or resale key.

• Hermes Skill: install API Relay Audit as a Hermes Agent skill and run the

same local 14-step LLM proxy security audit from an agent workflow.

These files do not certify that a relay is safe. They help agents generate a local, reviewable Markdown report before trusting a relay path.

Evidence Boundaries

Natural-language self-identification is treated as a consistency signal, not upstream proof. A response saying it is Qwen, DeepSeek, GPT, or Claude can indicate a mismatch, but it does not by itself prove that a provider substituted the upstream model.

Stronger claims require corroborating evidence such as raw response JSON, request IDs, provider/model metadata, stream signatures, transparent-log hashes, and reproducible runs. Public submissions should use redacted report artifacts and never include API keys, raw headers, full response bodies, wallet material, private relay traffic, or user data.

Web3 Wallet Safety Checks

With --profile web3 or --profile full, API Relay Audit adds wallet-oriented prompt injection probes inspired by signature-isolation risks:

• ETH transfer guidance checks
• Signed-transaction refusal checks
• Private-key leak refusal checks

These probes are model-agnostic, but they are intentionally profile-gated so general relay audits stay focused.

Working Model

your machine
  -> audit.py / scripts/audit.py
  -> chosen relay endpoint
  -> Markdown report + optional hash-only transparent log
  -> optional: redacted evidence issue for maintainer review

Community evidence is shape-checked by GitHub Actions, but publication still requires maintainer review. Operators keep a separate response path, and sensitive vulnerabilities belong in the disclosure path described in SECURITY.md.

Project Status

Metric · Current value Version · v2.5.0 Audit steps · 14 Risk matrix · 6D pytest collected tests · 885 CLI flags · 23 Runtime profiles · general, web3, full

Example Report And Live Page

• GitHub Pages: toby-bridges.github.io/api-relay-audit
• Chinese landing page: toby-bridges.github.io/api-relay-audit/zh/
• Example report: sanitized fixture report
• Guides:

AI API relay / LLM proxy, Claude relay audit, tool comparison, prompt injection in proxies, Web3 wallet prompt injection,

• Contributors / Credits: CONTRIBUTORS.md
• Security policy: SECURITY.md
• Contributing guide: CONTRIBUTING.md
• Social: X @li9292

FAQ

What is an API relay or LLM proxy?

An API relay or LLM proxy is a third-party service between you and an AI provider such as Anthropic or OpenAI. It forwards your requests upstream, but it can also inject hidden instructions, swap models, truncate context, rewrite tool output, or leak credentials in error responses.

Is it safe to enter my API key?

API Relay Audit runs locally, so your API key is sent only to the relay URL you specify. The standalone version is a single Python file with zero Python package dependencies, which makes it easier to inspect before running.

What does prompt injection mean here?

Prompt injection means the relay may prepend or insert hidden instructions into your request. API Relay Audit compares expected and actual token usage, tries prompt-extraction probes, and records evidence when the relay appears to add or reveal hidden prompt content.

What is model substitution?

Model substitution means the relay claims to provide one model but may expose evidence signals for another model identity, route, or upstream channel. API Relay Audit checks non-Claude identity patterns, anchor phrases, stream model identity, latency variance, and channel evidence where available; those signals require corroboration before making provider-level claims.

What is tool-call rewriting?

Tool-call rewriting means the relay modifies package-install commands or tool-like output in the model response. API Relay Audit sends pinned package commands and compares the returned text to detect proxy-layer supply-chain tampering.

What are SSE anomalies?

SSE anomalies are stream-level integrity issues in Anthropic-style streaming responses. API Relay Audit checks event types, usage monotonicity, thinking signatures, and stream model identity when the relay supports that format.

What Web3 wallet risks does it check?