Starting July 27, 2026, GitHub will cut public bug bounty payments at each severity level by at least half. Key findings will drop from $20,000-$30,000+ to a fixed $10,000, while its permanent invitation-only VIP tier will pay $30,000 or more.
Reports filed before that date, including reports already in GitHub’s growing triage queue, will retain the previous payment terms.
GitHub said the changes are aimed at reducing noise while giving established researchers faster responses, higher rewards, and closer access to its security engineering team.
“You cannot earn more by submitting more,” the company said. “You earn more by submitting better.”
Public program moving from flexible limits to fixed payments:
- Low: $250, below $617-$2,000
- Medium: Under $2,000, $4,000-$10,000
- High: Below $5,000, $10,000-$20,000
- Critical: Below $10,000, $20,000-$30,000+
Hacker News calculated that when measured against the lows of GitHub’s previous categories, the new public rates are 50% lower for medium, high, and critical findings, and about 59% lower for low-severity reports.
GitHub said fixed payments should remove uncertainty and triage overhead, though it can still give discretionary bonuses for exceptional work.
The VIP schedule sets payouts at $1,000 for low-severity findings, $7,500 for medium, $20,000 for high, and $30,000 or more for severe vulnerabilities.
Researchers can qualify for the private program by reporting at least one severe, two high, four moderate, or seven low-severity vulnerabilities. The announcement did not specify a deadline for meeting those criteria or say whether qualification guarantees an invitation. GitHub said the full criteria will appear on its public HackerOne program page.
GitHub has not disclosed the HackerOne signal limits it will enforce. The company says researchers below it will receive a maximum of four initial submissions.
Separately, HackerOne’s general rules give new researchers four test reports per program within a 30-day period.
When your own AI finds the bug first
GitHub’s report controls come as AI makes it cheaper to generate candidate findings and cheaper to repeat code reviews. More researchers can draw tentative conclusions, while internal teams can scan code, validate issues, and improve release and commit pipelines before external reports arrive.
A day before GitHub’s announcement, Google introduced Gemini 3.5 Flash Cyber, a lightweight model for finding, verifying, and patching software vulnerabilities. Google said the model will initially be available exclusively to governments and trusted partners through its code-protection agent, CodeMender, as part of a limited pilot.
Google said the model can be applied repeatedly to examine more code paths without using a larger frontier model for each attempt. The company deploys it for frequent repository scans, time-sensitive launch reviews, and commit-scanning pipelines.
In tests run by Google, Gemini 3.5 Flash Cyber found 55 unique confirmed V8 issues, compared to 47 for mainline Gemini 3.5 Flash and 36 for Cloud Opus 4.6.
Google separately said its cloud vulnerability research team used the model to find a remote code execution flaw in a public API and a memory-corruption flaw in a sensitive production service within two hours. The model then generated what Google described as a 100%-reliable RCE exploit that bypassed ASLR and W^X. Benchmark statistics and production exploitation results are reported by Google and have not been independently verified.
An internal security team can provide an agent repository reference, a project-specific threat model, and a validation environment tailored to the running system. Systems like OpenAI’s Codex Security can then test the findings, generate working proofs of concept, and propose solutions to the system’s intent and surrounding behavior.
Instead of waiting for scheduled evaluation or external reports, work can occur during development and at each relevant commit. AI doesn’t replace penetration testing, but it is becoming easier to automate source-code review, test generation, and first-pass validation.
Human testers retain greater value where the work requires exploiting vulnerabilities across trust boundaries, identifying business-logic failures, modeling realistic attack paths, and proving material impact.
Curl maintainer Daniel Steinberg terminated the project’s cached bug bounty at the end of January 2026 after its confirmed-vulnerability rate dropped below 5% amid a surge in AI-generated junk reports.
By April, when Curl ended the cash rewards and returned to HackerOne, reports were coming in at nearly double the rate of 2025 and 15–16% of vulnerabilities were confirmed. Steinberg said almost every report appears to be AI-supported and most are now of high quality.
Overall, GitHub’s reporting controls, Google’s repeated model calls, and Curl’s increasing submission volume point to the same change. AI may overwhelm maintainers with junk, but it can also make capable researchers faster and let internal teams check more code, more often.
The search for a plausible looking candidate is becoming abundant. Triage, exploit proof, product reference, disclosure and remediation remain disrupted. A credible exploit, a product-specific attack chain, or a discovery that crosses a threshold misunderstood by the vendor remains rare.
Signaling requirements and low public rewards can suppress automated noise, but they can also make entry difficult for capable researchers without an established HackerOne history. For a new HackerOne researcher, the four-report program limit leaves little room for mistakes, unfamiliarity with GitHub’s security model, or a legitimate finding that initially scores lower than expectations.
The invitation-only structure also concentrates GitHub’s closest researcher connections among those who have already been successful in the program. This can improve speed and report quality. It can also narrow down the scope of people vetting the platform, which is one of the main benefits of a public bounty program.
The restructuring follows a May 2026 policy change that called for closer attention to working proofs of concept, demonstrated impact, validation before submission, and GitHub’s scope and disqualification findings.
GitHub said it welcomes AI-assisted security research and already uses AI in its internal security programs. Researchers remain responsible for reproducing and verifying anything their instruments produce.
“Devices don’t matter,” GitHub said. “Quality of work matters.”
As of July 22, a review by The Hacker News found that GitHub’s rewards page still listed $20,000-$30,000+ for critical reports, while its FAQ retained the previous VIP eligibility test and two reports submitted during the past two years that earned at least $20,000. The FAQ also states that meeting those criteria does not guarantee an invitation and that GitHub reviews candidates quarterly.
