Munim Technologies

Research/Software

Fewer audit contests, more submissions, less money each

From early 2024 to early 2026, public smart-contract audit contests on Sherlock and Cantina fell 83% while the submissions they drew rose 24%, so each contest received 7.3 times as many; prize money per submission fell 81%, and the platforms' rules on AI-made findings range from an outright ban to admitting AI auditors as participants.

Nov 9, 2026 · 6 min read

Public audit contests pay security researchers to review a project's smart-contract code for a fixed time and split a prize pot among the valid findings. We built censuses of every public contest on six platforms from their own pages, APIs and repositories. On Sherlock and Cantina, contests that started in the first half of 2026 numbered 13, against 76 two years earlier, a fall of 83%. Yet the submissions those contests received rose 24%, from 15,787 to 19,614. Each contest now draws 7.3 times as many submissions, and the headline prize money behind each one fell 81%, from $451 to $85. For anyone deciding whether to take part, with or without AI tools, these are the numbers that set the odds.

76 → 13

public Sherlock and Cantina contests, 2024 H1 to 2026 H1

208 → 1,509

submissions per contest over the same period

$451 → $85

headline prize money per submission

The question

Researchers know contests have become scarce; Immunefi reported the fall in contest counts in September 2026. What we could not find published was the other side of the market: how many submissions the remaining contests receive, and how much money stands behind each one. We also wanted a dated record of each platform's rules on findings made with AI, since those rules decide who may compete at all.

What we did

We counted every public Sherlock contest with a submission count (259) and every ended public Cantina competition with one, and added them up by the half-year each contest started. The twelve Ethereum client competitions of February 2025 shared one pot, so they count as one contest. For HackenProof, the only platform that shows what each program actually paid next to its promised maximum, we recorded all 102 audit programs and fetched each program page, including any sentence about AI. Tests in the repository pin every number below to the data files.

What we found

Submissions per public contest, Sherlock and Cantina

By the half-year a contest started. Submissions include invalid, duplicate and spam reports. Source: cross_platform.py --table (data/sherlock_contests.csv, data/cantina_census.csv)

Measured from the 2025 H1 peak, contests fell 82% (71 to 13) but submissions only 41% (33,076 to 19,614). Submissions shrank far less than contests did, so they piled onto the few contests left.

Contests, submissions and prize money, Sherlock and Cantina

Started inContestsSubmissionsPer contestHeadline prize per submission (pooled)Median contest, prize per submission
2024 H17615,787208$450.85$260.25
2024 H26924,444354$456.42$209.17
2025 H17133,076466$315.04$129.20
2025 H23723,897646$211.75$92.41
2026 H11319,6141,509$85.25$41.54
Prize columns cover contests with a US-dollar pot only (for example 64 of the 69 contests in 2024 H2). Sources: cross_platform.py --table and prize_per_submission.py --table, both pinned to data/ by the repository's tests.

The fall is not an artefact of a few very large pots. Leaving out the nine contests with pots of $1 million or more, the pooled figure still falls 74%, and every half-year is lower than the one before. The median contest fell further, by 84%. In 2026 H1, 8 of 12 contests offered less than $100 of headline money per submission.

What was actually paid. Headline pots are not payouts. On HackenProof, 97 settled programs paid 22% of their combined maximums. Its DualDefense audits pay only for Critical findings, and 67 of 78 settled ones paid nothing. Grouped by start date, the money actually paid per submission fell 90% in a year.

HackenProof: money actually paid per submission

Settled programs only, by start half-year. Submissions rose from 940 to 3,588 a half-year while programs stayed at 19 to 25. Source: hackenproof_trends.py --table (data/hackenproof_programs_fetched.csv)

This has happened before. Code4rena went through the same crowding in 2022 and 2023: its median audit drew 29 High or Medium submissions in 2021 H1 and 356 in 2023 H2, and the share that became a distinct valid finding fell from 38.4% to 3.4%. What is new in 2025 and 2026 is that the supply of contests fell at the same time.

AI rules. HackenProof's program pages began carrying rules on AI-generated reports in July 2025. None of the 34 programs that started before then has one; 61 of the 62 that started from 14 August 2025 do, including all 43 that started in 2026. Across platforms, the rules vary far more than in general bug-bounty programs, where a published census of 53 programs found none that prohibits AI-written reports.

Platform rules on AI-made findings

PlatformStanceEarliest date we could verify
ImmunefiProhibited in its researcher terms; conditional on its rules pageBan announced Dec 2022
CodeHawksDiscouraged: risk of disqualification14 Sep 2023
Code4renaDiscouraged; from Sep 2025, findings of its owner's AI auditor treated as known issues25 Sep 2025 (V12 rule)
CantinaConditional: unvalidated AI findings can lead to a banUndated
HackenProofConditional, per program: PoC or human validation required16 Jul 2025
SherlockContests silent; Audit Engine admits AI auditors, per engagementUndated
Coded on Stingrai's published rubric for AI-report policies. Source: data/platform_ai_rules.csv; clauses read on 15-16 Oct 2026.

What it means

For a researcher, the typical public contest in 2026 offers about a sixth of the headline money per submission it did in early 2024, and on the one platform that publishes payouts, the money actually paid fell further still. Most of the AI rules are older than the collapse, so they cannot be read as a single response to it. Public counts also cannot say whether the extra submissions come from more people or from the same people sending more, for example with AI tools.

More research