0xSimao AI

0xsimao.com

Request an Audit

---

Accounting-first Solidity security audit skill, built from 0xSimao's 869 published findings (177 High, 247 Medium) across 143 reviews.

This does not replace a manual audit. Use it to catch issues early and to arrive at a review with a cleaner codebase. For a manual audit, contact 0xSimao on Telegram.

What it does differently

Most audit tooling scans for bad lines. This one writes down the protocol's accounting model first, then attacks it.

The ordering comes from the findings themselves. About a third of the High findings share one shape: value leaves the contract but the variable tracking it is never decremented, or it gets decremented in one branch of an if and not in the sibling branch. Early actors over-withdraw and whoever withdraws last cannot be paid. That bug is not visible line by line. It shows up once you have listed every function that writes every tracked total and asked who is left holding the loss.

How it runs

%%{init: {"flowchart": {"padding": 24, "nodeSpacing": 60, "rankSpacing": 55, "wrappingWidth": 900}}}%%
flowchart TD
  S["Source in scope"] --> O["Orchestrator"]
  O --> MM["<b>Money map</b><br/>• the assets, and where they sit<br/>• every total, and which functions write it<br/>• the invariants that must always hold<br/>• lifecycles: open, accrue, close, expire<br/>• actor cohorts, and what each one is owed"]
  O --> AT["<b>Asymmetry table</b><br/>• value moves, no total is written<br/>• written in one branch, not the sibling<br/>• never tracked anywhere at all<br/>• an operation with no inverse<br/>• a function in a family that is missing"]
  MM --> H["Both handed to every lens, with the source"]
  AT --> H
  H --> L["accounting-desync · share-exchange-rate · temporal-cohort · liquidation-solvency · cross-chain-state · rounding-precision<br/>ordering-mev · dos-griefing · access-trust · integration-assumptions · edge-states · flow-completeness"]
  L --> D["Dedup, then four judge gates"]
  D --> R["Severity, then the report"]

1. Money map. The orchestrator reads the source and writes out assets, tracked totals, the asymmetry table, invariants, lifecycles and actor cohorts. Around 200 lines, injected into every lens.
2. Twelve lenses in parallel. Each is a separate subagent with the full source, the money map, the method and its own specialty. No lens sees another lens's output, so two lenses landing on the same bug means something.
3. Dedup. Hard gates for function isolation, mechanism preservation, mitigation preservation and completeness.
4. Judge. Four gates, then severity calibration.
5. Report. Description and Recommended Mitigation per finding.

The lenses

#	Lens	Owns
1	accounting-desync	tracked totals drifting from reality, the largest class
2	share-exchange-rate	claims vs value, round-trip profit, redemption in terminal states
3	temporal-cohort	who gets the distribution, join-before / leave-before, index checkpoints
4	liquidation-solvency	health math, blocked liquidations, bad-debt clearing
5	cross-chain-state	LayerZero/CCIP global-state overwrite, debited-here-credited-nowhere
6	rounding-precision	direction, truncation, decimals, casts, overflow
7	ordering-mev	init races, unprotected protocol swaps, discrete-jump arbitrage
8	dos-griefing	unbounded loops, poisoned batches, pause interactions
9	access-trust	callbacks, approval abuse, unvalidated targets, composed privilege
10	integration-assumptions	token quirks, oracles, external protocols, chain environment
11	edge-states	zero, one, first, last, expired, paused, capped
12	flow-completeness	the gap hunter: missing calls, asymmetric branches, absent siblings

Install

Works with any coding agent that can run a shell, read files and spawn parallel subagents. SKILL.md is plain markdown with no vendor tool names or APIs in it.

Paste this into your agent, whichever one you use:

Install the 0xSimao AI audit skill:

1. git clone https://github.com/0xsimao/0xsimao-ai.git ~/0xsimao-ai
2. Work out where YOU load instructions from: your own skills folder,
   rules directory, or instructions file. If you have a skills or plugins
   directory, copy the clone into it under the name `0xsimao-ai`.
   Otherwise append to whichever instructions file you already read:

   When asked to run "0xSimao AI", "0xSimao audit", or to audit Solidity
   for security, read ~/0xsimao-ai/SKILL.md and follow it exactly.

3. Tell me where you installed it and how to invoke it from now on.

Your agent knows its own config layout, so it can pick the location without this README guessing at it. Ask for a user-level install to get the skill in every project, or project-level to scope it to one repo.

To update: run 0xSimao AI update, or cd ~/0xsimao-ai && git pull.

If that fails, or your agent has no plugin system, skip the install and say this in the repo you want audited:

Read ~/0xsimao-ai/SKILL.md and follow it to audit this codebase

Everything above only saves you from retyping that line. Pasting the contents of SKILL.md into the chat works as well.

Usage

run 0xSimao AI                      # full repo
run 0xSimao AI on Vault.sol         # specific files
run 0xSimao AI --file-output        # also write the report to disk

Runs twelve subagents in parallel over the in-scope source. Token spend is significant on a large codebase, so scope to specific files while iterating.

Agents that cannot spawn subagents fall back to running the twelve lenses one after another in a single context. Slower, and the lenses lose their independence, but the method still holds.

Benchmark

Sherlock's DODO Cross-Chain DEX contest, 1,632 nSLOC, 17 judged issues (5 High, 12 Medium). Audited at the judged commit with the findings held out of every agent's context.

	Recall	Runtime	Tokens
0xSimao AI	15/17 (88.2%)	~11 min	~1.1M
pashov solidity-auditor v3	14/17 (82.4%)	19–28 min	3.3–4.8M

This is not a head-to-head. The second row is there for scale. The two tools were run by different people at different times, each scored by its own author, with no shared harness. The claim being made is only that the skill finds real issues in a real contest at a useful rate.

Both numbers are self-reported. Recall is also the easier half of the problem, since neither figure says anything about precision. One run, one contest.

A rerun could invert the numbers. One run of a stochastic twelve-agent pipeline is a sample of size one. Run the same skill against the same commit again and it can miss what it caught and catch what it missed. The same applies to the other tool. Any single score here is evidence that the approach works, not a ranking.

The failure pattern is more interesting than the score. The run caught the two rarest issues in the set, submitted by 4 and by 7 of the 136 auditors credited in the judged report, and missed one that 13 auditors found. These tools do not fail uniformly. They fail unpredictably, which is why a quiet report is not the same thing as coverage.

Layout

SKILL.md                              orchestrator
references/
  simao-method.md                     the 7-phase method, how to think
  severity-calibration.md             four gates + severity assignment
  report-formatting.md                the report format
  attack-lenses/
    shared-rules.md                   output format + reasoning protocol
    <12 lens files>

Notes

Every attack surface in every lens is backed by verbatim titles from real findings. Keep that up when you extend a lens: add the pattern and the finding that proves it happens.