11 valid bugs in 20 live bounties and audit contests

Between March and November 2026 the QuillShield team, with QuillShield as its standard first pass, filed findings in 20 live Solidity bug bounties and audit contests. Judges confirmed 11 as valid, 4 of them filed as High, and Code4rena credits the team on a Medium that Intuition fixed. Another 18 are in judging.

Published . 7 minute read. By the QuillShield team.

We wanted a test of QuillShield that we could not grade ourselves, and live bug bounties and audit contests are that test. Other researchers hunt the same code at the same time, and a judge who does not work for us decides what counts. From March to November 2026 our team's practice was to start each program with QuillShield and file what held up under human review. This post covers what we found and what the judges confirmed.

What we did and why

Our EVMbench run measures how many known bugs QuillShield finds, and we graded it ourselves. In a bounty or a contest nobody knows the bugs in advance, and the platform or the protocol's team rules on each report.

Between March and November 2026 the QuillShield team filed 43 findings in 20 Solidity programs: 7 audit contests and 13 bug bounties.

How each run worked

The team's practice was to run a full QuillShield audit of the in-scope code before anyone reviewed it by hand. That was already how it worked in March, when QuillShield's report on the Intuition contest came back with proofs of concept. In July the team made it a rule, with manual review limited to two or three contracts out of ten. The same month it started skipping projects that top audit firms had already reviewed heavily.

Auditors then checked each QuillShield finding against the program's known issues and earlier audits, and sorted it into real, false, already known, or needs a proof of concept. The team's written process says nothing goes out until a person has reproduced it with a failing test or a proof of concept. The team also used other AI workflows and skills, fuzzing and manual review, and later a tool that maps a protocol's external dependencies. So every result below is the team's.

One staking protocol's bug bounty shows the effort. QuillShield reported eight findings and the team filed one, which was judged valid. The engagement log records 25 hours, split into 10 to understand the protocol, 2 for setup, 3 to triage QuillShield's output, 3 of manual review, 4 for the proof of concept and 3 for the report.

What the judges confirmed

Judges confirmed 11 findings as valid: 4 filed as High, 6 as Medium and 1 as Low. All 11 were duplicates of bugs other researchers also reported, and none was judged unique. Another 18, filed as 4 High and 14 Medium, are in judging in one audit competition, with results due about 20 November. We will update this post when they are out.

Where every finding stands Findings Filed as High Medium Low Not recorded
Judged valid 11 4 6 1 0
In judging 18 4 14 0 0
Not accepted or out of scope 9 2 3 1 3
No ruling on record 5 0 3 0 2
All findings 43 10 26 2 5

Severity is as filed. Where our records disagree, on one finding's severity and on the count filed to four bounties, we took the less favourable reading. Not counted are 14 findings in five programs written mostly in Move, Go or Rust, which QuillShield does not review, and two tracker entries that log submissions without a count.

Programs we can show

We name a contest only when its results and report are public. We describe a program without naming it when the result reveals nothing anyone could exploit. Everything else stays in the totals, and that keeps some of our strongest results unnamed: three of the four valid findings filed as High are on programs whose rules bar discussion.

Program Type Month Bug class Filed as Ruling What it shows
Intuition on Code4rena Contest Mar 2026 Reward accounting at an epoch boundary Medium, plus one Low Medium judged valid as M-02, then fixed QuillShield's report came with proofs of concept
A staking protocol Bug bounty Jul 2026 Fee accounting (griefing) Medium Judged valid QuillShield flagged eight candidates, and one became a valid Medium
A token launch protocol Bug bounty Aug 2026 Theft of funds High Judged valid, as Medium The program rated it a step lower than we did
A tokenised asset protocol Bug bounty Apr 2026 Input validation Medium (one record says High) Judged valid We count the lower of our two severity records
A DeFi automation protocol Bug bounty May 2026 Denial of service (griefing) Medium Judged valid A denial-of-service path the program confirmed
A lending and automation protocol Contest Oct 2026 18 findings across classes 4 High, 14 Medium In judging The first test of the dependency map and the team's newer workflows
A stablecoin protocol Bug bounty Aug 2026 Not recorded Not recorded Closed, no ruling on record QuillShield flagged 1 High and 22 Medium candidates for triage
Metric on Sherlock Contest Jul 2026 Share valuation Medium No ruling on record 77 contracts in scope, all run through QuillShield

Intuition on Code4rena: a Medium the protocol fixed

Code4rena ran the Intuition audit from 4 to 9 March 2026, on five contracts and 844 lines of Solidity. QuillShield's report came with proofs of concept, and our auditors checked each finding. One candidate said a governance setting could change reward maths partway through an epoch. Asked for a proof of concept, QuillShield's validation argued against its own finding, because the setting belongs to a trusted role behind a three-day timelock. The auditor moved on.

What we filed was judged a Medium. Checkpoints made at an epoch boundary could qualify for the previous epoch's rewards. The published report lists it as M-02 and credits QuillAudits_Team among the wardens who reported it. The contest had two Mediums and no Highs. Intuition fixed both, and a mitigation review confirmed the fixes.

What we built next

Each program sharpened the method. The team started building about 100 bug patterns for each type of protocol, a map of each protocol's external dependencies and a shared list of invariants. The dependency map went into our entry in the competition now in judging, where we filed 18 findings, 4 of them as High.

What this means before your own bounty or contest

Every valid finding in this post was a real bug in live code, and each was there to be caught before the program opened. That is the cheapest moment to find one. Run an AI reviewer before you open a bounty or book a contest, fix what it finds, and let the crowd spend its time on bugs that need a person. Three habits from these programs help any team:

  • Treat severity as a starting guess. One High we filed was judged Medium.
  • Learn what the protocol calls intended. Code that looks wrong can be a design choice, and on Intuition QuillShield's own validation argued one such candidate down before anyone filed it.
  • Read the program's rules before spending the hours, since scope and the minimum severity decide what can be filed.

The benchmark points the same way. On EVMbench, a set of known bugs from past audits, the QuillShield alpha found 98 of 117 (83.8%) in our own run, as the EVMbench post explains.

What to do next

If you are about to open a bounty or book a contest, run QuillShield on the in-scope code first and fix what it finds. Start an audit, or follow the guide to your first audit.

Common questions

What did QuillShield find in live bug bounties and audit contests?

Between March and November 2026 the QuillShield team filed 43 findings in 20 live Solidity bug bounties and audit contests, with a full QuillShield audit as its standard first pass. Judges confirmed 11 as valid, 4 of them filed as High, and 18 more are in judging. All 11 were duplicates of bugs other researchers also reported, and none was judged unique.

Which result can I check in public?

Code4rena's Intuition report lists the team's Medium as M-02 and credits QuillAudits_Team among the wardens who reported it. Intuition fixed it, and a mitigation review confirmed the fix.

How did QuillShield fit into each submission?

The team's practice was to run a full QuillShield audit of the in-scope Solidity code before manual review. People then triaged its output, reproduced each finding and wrote it up, and they also used other AI workflows and skills, fuzzing and manual review. The records name QuillShield as the source for some findings, including the Intuition Medium and a staking protocol's Medium.

Why are most of the programs not named?

Bug bounty and contest platforms limit what researchers may say about reports. We name a contest only when its results and report are public, describe a bounty without naming it when the result reveals nothing exploitable, and keep everything else in the totals.

Sources

QuillShield is building the security layer between AI and money, so you can keep shipping between audits.

Sign-up is free, and new accounts get $50 in credit, good for up to half of an order.

  1. 1Sign up and import a repo
  2. 2See the exact price before anything is charged
  3. 3Approve it, and the audit starts