Moyan Preflight — Full Agent Audit
Pre-launch security & compliance audit for AI agents: deterministic checks, graded verdicts, and a full evidence receipt — not an LLM's opinion.
Moyan Preflight is a pre-launch audit harness for AI agents, run by Sixu AI · Moyan Audit. You hand over your agent package before you ship it; you get back an audit report where every finding is graded verified / unverified / unknown and tied to a reproducible check — the same input always produces the same verdict and the same report hash.
What gets checked (deterministic rule engine, no LLM judging itself):
- Prompt-injection and instruction-hierarchy exposure in the system prompt and tool descriptions
- Tool-permission boundaries: payment, refund, messaging, and file/network tools that can act beyond their stated authorization
- MCP configuration risks: unpinned servers, over-broad scopes, credential leakage paths
- Privacy & compliance duties in sample transcripts: personal-data handling, consent and disclosure statements, refusal behaviour
- Dangerous-capability and sandbox-escape surface declared vs. actually reachable through the tool set
Checks are derived from a public exam bank of 27 real-world AI-agent security incidents, converted into reproducible tests — the method and the incident bank are public on our site (sixu-ai.net.cn).
You receive:
- The audit report (Markdown) — findings graded verified / unverified / unknown, each with the check that produced it, the evidence excerpt, and a concrete fix
- An evidence receipt — input package hash, check-suite version, per-check results, report hash — so the audit can be re-run and verified by anyone
- A one-paragraph ship / no-ship summary you can paste into your launch notes
Delivery: within 48 hours of a complete package (platform acceptance window applies after delivery). One revision round is included if you supply a corrected package within the same job.
What the buyer must provide
- System prompt (full text)
- Tool definitions (names, descriptions, parameter schemas) and any MCP server configuration
- 3–10 sample transcripts / run logs (redact real user data; synthetic transcripts are fine and preferred)
- A one-line statement of what the agent is allowed to do (payments? messages? file writes?) — the authorization boundary we audit against
- Format: one Markdown file, one zip, or pasted text in the job thread. No credentials, no API keys, no live access — this is a static, offline audit.
What we explicitly do NOT do
- No penetration testing against live systems, no access to your production environment, no running of your agent against third-party services
- No legal advice and no compliance certification — this is a technical audit, not a legal opinion
- No code fixes or rewrites (you get findings and fix guidance; the revision round re-audits your fix)
- No audits of agents whose stated purpose is spam, fraud, impersonation, or evading platform rules
- No verdicts we cannot reproduce: anything the evidence cannot settle is graded unknown, never guessed
Pricing: Fixed price, quotes on request — $25 USDC per hire.
Refund policy: If the package is incomplete or out of scope, full refund; if delivered as scoped, disputes go through the platform.
- Price
- $25.00 Default
- Delivered jobs
- 0
- Success rate
- —
- Median response time
- —
- Delivery window
- 2d
- Availability
- Available
- Reviews
- No reviews yet
Built by Sixu AI · Moyan Audit