Proof for every byte.
Your code, your installs, your agents: a receipt anyone can check, instead of a badge you have to trust.
Your cursor is the verifier. Every byte it looks at gets checked; one random byte is spot-checked.
Old code, new exploits.
The worst breaches hit code everyone already trusted. Mature, widely used, reviewed many times over, and still carrying a flaw for years, until an attacker found it first. Age and reputation are not evidence. A check you can repeat at every commit is.
- Shellshock · BashA command-injection flaw sat in Bash for about 25 years; mass scanning began within hours of disclosure.
- Heartbleed · OpenSSLA missing bounds check in the web's main TLS library leaked server memory, private keys included, for two years.
- Log4Shell · Log4jA lookup feature added in 2013 let a single log line run remote code, in software nearly every Java team ran.
- PwnKit · polkitAny local user could become root on most Linux systems, through a bug present since pkexec's first release in 2009.
- regreSSHion · OpenSSHA bug fixed in 2006 came back in a 2020 release and opened unauthenticated root code execution on glibc Linux.
- Wormhole · SolanaThe bridge checked signatures through a deprecated Solana function; an attacker minted 120,000 wETH, about $320M.
Lock the inputs before anyone sees the answer.
The commit, the files, the audit spec, the model id. Hashed into one commitment and written to the log first. From here nothing can quietly change.
Tie the dice to a block nobody has seen yet.
The seed comes from the commitment and a future Solana block. Neither the customer nor ByteGuard can re-run until the answer is the one they want. One run per commit. No re-rolls.
Record the answer, word for word.
Today a hosted model does the audit. Its output is hashed and committed as it comes; the verdict follows from it deterministically.
We do not claim byte-identical re-runs today. That is the Phase 4 upgrade, on a self-hosted deterministic node.
{
"summary": "The escrow program has one high-severity authorisation gap and two lower-severity issues. None are critical.",
"findings": [
{ "severity": "high", "category": "signer check",
"title": "Escrow authority is not required to sign on cancel",
"file": "src/lib.rs", "line_start": 142, "line_end": 158,
"fix": "Mark initializer as Signer and add has_one = initializer." },
{ "severity": "medium", "category": "arithmetic",
"title": "Unchecked multiplication when computing taker amount",
"file": "src/lib.rs", "line_start": 96, "line_end": 99,
"fix": "Use checked_mul and return a custom error on overflow." },
{ "severity": "low", "category": "closed-account reuse",
"title": "Vault closed without zeroing discriminator",
"file": "src/lib.rs", "line_start": 171, "line_end": 176,
"fix": "Use Anchor's close constraint." }
]
}Sign it into a log nothing leaves.
Every entry gets a one-time hash-based signature (LMS, RFC 8554). Entries hash up a Merkle tree. The root is signed every hour. Daily roots anchored on Solana come in Phase 2.
Check it yourself. In your browser.
Twelve checks, none taken on trust from the server. Signatures, log inclusion, the recomputed verdict, the Solana block.
- 1Receipt keypass
- 2Entry signaturespass
- 3One-time keys used oncepass
- 4Entry orderpass
- 5Audit specpass
- 6Source filespass
- 7Seedpass
- 8Model outputpass
- 9Verdictpass
- 10Signed tree headpass
- 11Log inclusionpass
- 12Solana blockpass
Two tools live, one in build. Four on the roadmap.
One engine behind all of them. We announce only what is live.
Scanner
LiveAn instant risk scan of a Solana token or program. Thirteen checks, a verdict, a badge. Free and anonymous.
About Scanner →Shield
LiveSee who may move your tokens, and revoke that access in one click. Your wallet signs; ByteGuard never holds a key.
About Shield →Audit
In buildAn AI security audit of a Solana program at a git commit, on a locked seed, with a signed receipt anyone can verify. One run per commit, no re-rolls.
About Audit →Scan something.
A token mint or a program id. No wallet, no account, about a second.