GitHub has substantially redesigned its bug bounty program for reports submitted from July 27, 2026. The platform now distinguishes between a public program, available under reputation requirements, and an invitation-only VIP tier offering the largest rewards. This is not simply a change to the payout table. It addresses an operational problem: an influx of weak, verbose or AI-assisted reports submitted without sufficient validation.
In an earlier post, GitHub said that a strong report needs a short summary, reproducible steps backed by evidence and a concrete explanation of impact. Long theoretical narratives and AI-generated filler have the opposite effect when the actual vulnerability is difficult to find. The new structure therefore attempts to shift incentives from the number of tickets toward their quality.
The decision matters beyond GitHub. It shows how generative AI is changing the economics of vulnerability disclosure for both researchers and product security teams.
Two tiers for two levels of established trust
The public program remains an entry point, but it now uses fixed awards. GitHub lists $250 for a low-severity vulnerability, $2,000 for medium, $5,000 for high and $10,000 for critical findings. Researchers invited to the VIP program can receive roughly three to four times those amounts depending on severity.
That difference rewards more than a vulnerability. It rewards an established working relationship: the ability to validate a result, communicate clearly, follow program rules and cooperate with the team during remediation. The public tier can surface new talent, while the VIP tier gives GitHub a smaller group whose reporting signal is already understood.
New participants also need a minimum level of reputation on HackerOne to enter the public program. GitHub provides a limited number of opportunities to establish that credibility. The filter may reduce noise, but it also raises the barrier for a capable researcher who has not yet built a history on the platform.
AI lowers the cost of a report, not the cost of evidence
A language model can turn rough notes into structured prose, explain a vulnerability class or suggest reproduction steps. Used carefully, it can help create a clearer report, particularly when a researcher writes in a second language. It may also draft an initial script or summarize large logs.
The problem starts when generation replaces verification. A model can invent a function, infer an impossible impact or present unusual but harmless behavior as exploitable. The resulting document often looks persuasive, with precise vocabulary, neat sections and familiar remediation advice. Triage still has to rebuild the entire analysis if no request, trace or demonstration supports the conclusion.
AI therefore reduces the cost of producing a ticket but not the cost of proving it. It may transfer that cost from the submitter to the security team. When thousands of people can rapidly submit hypotheses, the reward system must draw a sharper line between automated exploration and a validated discovery.
What a high-quality report must demonstrate
The first requirement remains reproducibility. A report should specify prerequisites, the account used, configuration, requests sent and observed result. A screenshot can show a symptom without establishing its cause. A minimal script, sanitized HTTP exchange and test environment allow triage to confirm behavior more quickly.
The second requirement is impact. An error exposing information that is already public is not equivalent to cross-account data access. Researchers should describe what a realistic attacker can achieve, which permissions are required and which limits they encountered. Inflating severity damages credibility even when the underlying bug is real.
The third is concision. Relevant context should remain, but ten generic paragraphs about SQL injection add nothing when the product does not use SQL at that point. AI assistance should remove ambiguity rather than inflate the document.
Finally, claims must be separated from hypotheses. A statement that a behavior "could allow" an outcome requires another experiment. Labeling that uncertainty helps the team decide whether to investigate without confusing a possibility with demonstrated exploitation.
Triage teams need to evolve as well
Reputation filtering will not solve the problem by itself. A modern team needs ways to group duplicates, identify nearly identical descriptions and quickly extract verifiable elements from a long report. AI can assist on this side too, provided it supports an analyst instead of automatically closing tickets.
Submission forms can require structured fields for the affected asset, prerequisites, expected result, actual result, evidence and impact. A validator can confirm that an attachment exists or that a scenario contains steps without pretending to know whether the vulnerability is genuine. Fast responses also matter: a serious researcher left for weeks in a saturated queue may move to another program.
GitHub says it is investing in faster response times, clearer severity reasoning and greater community engagement. The financial table is only one part of the answer. Bug bounty quality also depends on how valid submissions are received and handled.
The risk of making a program too closed
A VIP tier improves average signal, but it creates a concentration risk. Regular researchers learn the methods a company values and may explore similar areas. A newcomer, academic or specialist in an unusual technology can contribute a different approach precisely because they are outside the established circle.
If public rewards become too low relative to the work required, some vulnerabilities may be sold to brokers, disclosed elsewhere or simply abandoned. GitHub therefore needs a legible path toward VIP status and useful explanations for rejected reports. The public program cannot become merely a costly filter paid for by researchers.
It is also important not to equate AI use with poor research. A model-assisted report can be excellent when it includes reproducible evidence and its author understands every claim. A fully human report can still be vague. The relevant criterion is the quality of the result, not the tool used to write it.
A new discipline for security researchers
For bug hunters, volume is becoming a less viable strategy. Before submitting, researchers should reproduce the issue on an account they control, reduce the scenario to its minimum, verify scope and retain evidence. Any AI-produced output should be reviewed like code from an untrusted source.
Models are most useful around the research process: preparing a checklist, rewriting an explanation, comparing behavior with a specification or suggesting edge cases. The decision to claim that a vulnerability exists must remain grounded in observation.
GitHub's redesign does not mark the end of open bug bounty programs. It shows that their scarce resource is no longer the ability to receive ideas, but the human time required to separate a real flaw from plausible prose. In that new balance, evidence is more valuable than ever.
