Claims lie. Code doesn't.

BYTEJURY: AI cybersecurityaudits by AI agents.Delivered by a jury.

Five AI agents audit the code behind a project: the token and every contract around it, from proxies and vaults to fee burners, Uniswap v4 hooks and bonding curves. They map who can mint, pause, blacklist, change fees or upgrade, hold every claim on the site and X up against the code, and cite a line, a transaction or a test for each finding.

Robinhood Chain, Ethereum, Base, Arbitrum and Solana. Tokens and any other contract.

Run an AI audit Browse audits

55 reports published.Pay per audit in $BYTE.Once anyone pays, the report is public and free.

Example verdicts (made-up tokens)

$PHOOD

High risk31/100$MOONCAT

Scam4/100$DUCKSWAP

Caution58/100$PIXELBANK

Safe86/100$GRUMPY

Caution52/100$TOASTER

Safe79/100

An AI audit of the whole system.Not just the token.

Websites and posts can say anything. The contract can't. The AI auditor reads the contract you paste and every contract it touches, writes a plain-English map of who controls what and where the money flows, and lists what it could not verify instead of guessing.

Source or bytecode
Verified source, read with line numbers. Not verified? The bytecode is decompiled and audited anyway.
Every related contract
Proxies and implementations, vaults, staking, fee burners and splitters, Uniswap v4 hooks, bonding curves, factories and registries.
Privilege map
Who can mint, pause, blacklist, change fees or upgrade, within which limits, and whether a timelock or a multisig stands in the way.
Money flows
Buy, sell, deposit, withdraw, claim and fee paths, step by step through each contract, and how each path could be abused.
Claims vs code
Every promise on the website and X is held up against the code and marked confirmed, contradicted or unverifiable.
Any contract
Tokens, routers, vaults, multisigs, timelocks, v4 hooks. Paste an address, a website, an X handle or an explorer link.

One finding, with its receipt.

Every related contract, not just the token. The example reads 6 of them.
Each privileged function gets a row: who may call it, what it changes, its live value and its risk.
Each finding points to a line of code, a transaction or a test result.

Example report. PepeHood $PHOOD is a made-up token; names, addresses and numbers are invented to show what a report looks like. The flagged function is setBurnToken(address) in HoodBurner. See real checks →

HoodBurner.sol — contract.exe

EXAMPLE

31contract HoodBurner is Ownable {
32 IERC20 public burnToken;
33 address constant DEAD = 0x…dEaD;
34
35 /// Swap the collected fees into burnToken
36 /// and send the output to the dead address.
37 function burn(uint256 minOut) external onlyKeeper {
38 uint256 out = router.swap(fees(), burnToken, minOut);
39 burnToken.transfer(DEAD, out);
40 }
41 function setBurnToken(address token) external onlyOwner { (flagged line)
42 require(token != address(0), "zero");
43 burnToken = IERC20(token);
44 emit BurnTokenSet(token);
45 }
46}

highOwner can switch the "buy-and-burn" from $PHOOD to any other token

HoodBurner · setBurnToken()

Function facts for setBurnToken(address) in HoodBurner
Signature	setBurnToken(address)
Access	owner (0xDEMO…0WNR, one wallet)
Effect	Switches which token the collected fees buy and burn. Any time, no timelock.
Live value	burnToken = 0xDEMO…C0DE (currently $PHOOD)
Location	HoodBurner.sol:L41-L47
Risk	high

Buy on fork
Sell on fork
Token tax 4%
Second wallet sells
Sell after +1h

Done. 6 related contracts read, 9 findings in the report.

Six AI agents. One verdict.

Each AI agent has one job, like a courtroom. The site analyst lists every claim the project makes, the auditor reads the code, OSINT checks X, the prosecutor argues the case against the project, and the trend scout reads the narrative on X for a separate heat score that never touches safety. Then the models review each other's findings as critics, and the judge weighs it all into a safety score from 0 to 100, where higher is safer.

Site analyst
Gemini 3.8 Flash
OSINT
Grok 4.7
Auditor
Claude Opus 5.5
Prosecutor
GPT-5.6 Sol
Judge
Claude Opus 5.5
Trend scout
Grok 4.7 · xAI

+ 3 critics: GPT-5.6 Sol, Claude Opus 5.5, Grok 4.7

deliberation.log — PepeHood $PHOOD

EXAMPLE

Site analyst · Gemini 3.8 Flash
Listed 6 claims the project makes on its site and X, to test against the code.
OSINT · Grok 4.7
The deployer funded 11 of the first 20 buyers an hour before launch.
high · confirmed by 3 models
Auditor · Claude Opus 5.5
Owner can switch the "buy-and-burn" from $PHOOD to any other token.
high · confirmed by 3 models
Prosecutor · GPT-5.6 Sol
2 of 6 claims contradicted, e.g. “100% of fees burn $PHOOD forever”.
Critics · GPT-5.6 Sol, Claude Opus 5.5, Grok 4.7
Cross-checked 7 AI findings: 1 disputed and dropped.
Judge · Claude Opus 5.5
Trades fine at 4% tax, but one wallet controls the burn and insiders bought the launch.
HIGH RISK · 31/100

PepeHood $PHOOD · 8 model runs · example

verdict.txt

findings: 9
claims checked: 6
contradicted: 2

HIGH RISK

31/100

Contract
38
Token
22
Social
35
Website
30

Higher is safer. Weights: contract 45%, token 30%, social 15%, website 10%. The token part was capped at 25: insiders bought the launch.

Example report · invented numbers

Example report. PepeHood $PHOOD is a made-up token; names, addresses and numbers are invented to show what a report looks like. The roles and the models are the ones a real scan uses. See real checks →

Evidence first.Then the AI.

An AI audit is only worth the evidence behind it. Everything a model cites is checked against the data we actually collected, and hard facts from the code and the fork set floors no model can argue its way past.

A finding without valid evidence (a code line, a transaction, a test result, a post) is dropped before the verdict.
Critics confirm or dispute every AI finding. One disputed more often than confirmed doesn't make the report.
A failed sell, an uncapped mint or a heavily sniped launch caps the score, however the agents argue.
Re-scans re-check every earlier finding against fresh data and show what was fixed, what is new and what got worse.
Reports are pinned to IPFS and anchored on Robinhood Chain: an ERC-7512 audit record for EVM contracts and an ERC-8004 validation from our registered agent.

floors.rules — applied before the judge

Examples of deterministic floors and what they do to the score
Hard fact	Effect
Sell reverts on the fork (honeypot)	Verdict SCAMoverall score ≤ 10
Uncapped mint held by a live owner	At least HIGH RISKcontract part ≤ 35
Upgradeable proxy run by one wallet	At least HIGH RISKcontract part ≤ 40
Heavily sniped launch	token part capped
Source code not verified	At least CAUTIONcontract part ≤ 60

AI findings can make a score worse. They can't lift it past a floor.

pepehood.example — test browser

EXAMPLE

https://pepehood.example/stake

REC4/4

unstake 4/4Done: 900 of 1,000 PHOOD returned

unstake→ HoodStakingwallet +900 PHOOD (10% exit fee)

Example run on a local fork: stake passed, unstake failed.

Then we test it live, with fake money.

On a private fork of the chain, with test ETH, we buy and sell the token, sell from a second wallet and sell again an hour later. A test wallet then opens the project's site and does what you would do: connect, approve, deposit, withdraw. No real funds move, and every step is screenshotted and recorded on video.

stake: 2 tx, all succeededpass
approve: exact amount, not unlimitedpass
unstake returned 900 of 1,000 PHOOD (90%)flagged
site says "no exit fee"; the contract took 10%flagged
RPC calls that reached a real node: 0pass

Example report. PepeHood $PHOOD is a made-up token; names, addresses and numbers are invented to show what a report looks like. In this example unstaking returned 900 of 1,000 PHOOD, although the site promises no exit fee. See real checks →

Plus, for tokens

Launch forensics.

Alongside the code audit we replay the launch block by block: snipers, bundles that share one funder, what the deployer did, and whether liquidity is locked or burned. A heavily sniped launch caps the token part of the score, however clean the code is.

Example report. PepeHood $PHOOD is a made-up token; names, addresses and numbers are invented to show what a report looks like. Its LP is locked for 90 days; its launch was sniped. See real checks →

launch.forensics — PepeHood $PHOOD

EXAMPLE

20 sniper wallets, first 12 seconds (of supply)38.4%
One funder, 11-wallet bundle (of supply)24.1%
Deployer sold (of own tokens)62%
Snipers still holding today (of supply)9.3%
Main pool LP locked (90 days) (of LP)100%

block 111 wallets, 1 funder, 24.1% in one blockdeployerbought 3% at launch, sold 62%ruletoken score capped at 25: insiders bought the launch (scored 22)

Funding traced for 20 of 20 early wallets

byte.split — where a scan fee goes

1 scan fee in $BYTE

…

burned forever

…

to the treasury

Free right now.Public forever.

When paid scans return, they cost $BYTE. Part of every fee is burned on-chain, the rest funds the treasury.
New audits are free right now, no wallet needed. Every report is public for everyone, saved to IPFS and anchored on-chain.
Holders scan free. Thresholds and free-scan allowance are set on-chain.

One scan: Reading chain…
Free scans: Reading chain…

Read live from the ScanPayments contract (scanPrice, getTiers, creditMode).