Paste a finding
The write-up you are about to submit, with its GitHub links to the contract code.
“deposit() mints shares before…”
Paste an audit finding. Basanos checks that its code links are real, reads the whole codebase and the contest's rules, and suggests the checks the finding still needs, each tied to what acceptance depends on. When it is not sure, it says so and leaves the call to you, with what it found.
Live demo on a closed Code4rena contest (Karak, July 2024). Measured on a second contest that was held out of development. Results, including what failed →
Audit contests pay bug hunters for findings a judge accepts. Many rejections are checkable before submission: a wrong line, a known issue, a missing proof. Basanos checks for those before you hit submit.
The write-up you are about to submit, with its GitHub links to the contract code.
Are the cited lines real and relevant? Is it a known issue? Does the code actually allow the attack? Is there proof?
Suggested checks, a verdict, and whether it needs your review. Every model “no”, anything uncertain, and anything where its own two layers disagree is left to you. A broken code link is flagged automatically, as a link to repair.
Built for bug hunters and the agents that hunt for them, and for contest platforms that triage thousands of submissions. See the full pipeline →
Pick an example, or paste a finding for the closed Karak contest. The examples include a real awarded finding, a known issue, a fake citation and an injection attempt.
Like an assayer rubbing gold on black stone, Basanos leaves a streak: its colour is the verdict, and underneath is everything that would make the finding stronger.
Basanos asks three times. It gives automatic advice only when all three agree, no code rule had to overrule a model, and the advice is not a “no”. Everything else needs your review. Every verdict is advice, not a guarantee.
The code supports the claim, the proof is in the finding, and it is not a known issue.
Your call; it has not yet said this of a real finding.Plausible, but something is still unshown. The verdict lists the suggested checks.
Add the listed proof.It believes something in the audited code stops the attack, and names the line. Left to your review.
Re-check that line.It quotes a known issue, bot finding or scope rule it believes applies. Code checks the quote exists, not that it applies.
Check the quoted rule yourself.No link points at audited code the finding actually talks about. Caught by code, at no cost.
Fix the links.The text tried to instruct the reviewer, or a screen held it. Nothing was decided.
Read it yourself.These 2024 findings could be in a model's training data. So we fixed each award-winning bug on a branch and ran the same finding again. The verdict moved with the code, which shows it reads the source. It does not prove the models have never seen these contests.
runs said “blocked by code” on the original, buggy code
said it once the bug was fixed: the answer moved with the code
times it pointed at blocking code, it was in the file the fix changed
injection attacks written by an outsider got an automatic “ready”
Contest judges triage thousands of findings. Some cite a line that says something else, repeat what the bot report already caught, or claim a loss nobody showed. Each of those costs a judge time and a hunter the payout. The fix is a touchstone, not another model.
Four live attacks with the expected result written down first, and the rubric mapped to evidence.
→ 02 · ProofEvery claim, checkedThe ledger and the finding tray recompute in your browser; the rest reproduces with npm run verify, including the results that went against us.
→ 03 · How it worksThe Touchstone PassWhat code decides, what Dolphin reads, what SERV guards, and what a check costs.
→ 04 · For agentsMCP, CLI and APIGive a bug-hunting agent one tool to call before it submits anything.
→