Why the latest Intigriti platform data challenges the notion that AI could replace bug bounty.
How researchers are moving from single prompts to multi-agent, always-on workflows.
Why AI raises the baseline for vulnerability discovery while the crowd expands its upper limit.
What CrowdRecon is teaching us about the coverage already created by hacker reconnaissance.
Why coordination, validation, and prioritisation become more important as both volume and quality rise.
In Common AI misconceptions debugged, I suggested that AI was increasing throughput rather than breaking bug bounty. Six months later, the picture has moved on. Submission volume is still growing, but validity has recovered sharply. High-severity findings are increasing. Repeat researchers are returning. The community is not simply producing more, but it is learning how to produce better.
AI is already automating more of the work and is accelerating reconnaissance, helping researchers iterate on payloads, generating proofs of concept, and making security testing more accessible. The floor is rising, but that misses the more useful question: what happens when thousands of hackers get access to the same technology and adapt it to their own methods, experience, and curiosity?
The answer is starting to show up in our Intigriti data. AI is not flattening the value of the crowd; it is giving the crowd more leverage. Researchers are combining models, building specialised agents, and monitoring programs continuously. They are using automation to cover more ground, then applying human judgment to find the application-specific weaknesses that modernised tools still struggle to see.
At the beginning of 2026, researcher engagement grew by roughly 320–330% over three months. AI was not the only cause, but it was clearly part of the environment around that change. New tools lowered the friction involved in learning, testing, and reporting, and the immediate effect was mixed. Platform validity fell to around 35%, and the rate among new researchers dropped as low as 18%. People could move faster before developing the judgment that comes with experience.
If we had stopped the analysis there, the pessimistic interpretation would have looked reasonable: AI increases participation, participation increases noise, and quality deteriorates.
That is not what happened next. By September 2026, the average validity ratio had climbed to a platform high of 68%. It increased by 23% in a single month while submission volume remained high. Submissions themselves grew by a further 12.5% month-on-month. Active researcher numbers rose by 15% in August and another 2% in September, without the same dilution in engagement quality.
The shift is also visible in outcomes:
High-severity vulnerability discovery increased by 12.9% in September 2026.
Payout per researcher increased by 5% in the same month.
Repeat submissions grew, suggesting researchers were finding programs worth returning to.
AI-specific vulnerability reports, including validated prompt injection, continued to rise sharply.
The data does not prove that AI caused the recovery. It shows that engagement and volume remained high while the proportion of valid findings improved significantly.
The reasonable inference, supported by what researchers tell us about their workflows, is that the community is becoming more sophisticated in how it uses AI. The first phase was about speed. The next phase is about orchestration, specialisation, and judgment. This is an important distinction. A tool can raise the floor quickly, and raising the ceiling requires people to learn its limits and combine it with other methods.
The phrase “AI-assisted hacking” can create the wrong mental model. It suggests a researcher asks one model to find a vulnerability and receives an answer. The reality is becoming much more operational. Researchers are combining multiple models and agents in parallel. One agent might map an attack surface. Another can investigate authentication flows. Others test payloads, analyse source code, operate a browser, or correlate results across tools. Open-source frameworks increasingly connect these agents directly to established security tooling through standard interfaces such as Model Context Protocol.
The strongest workflows also demand proof. An issue remains unconfirmed until an agent or researcher reproduces it with a working proof-of-concept. That moves the work from AI-generated speculation towards validated impact.
We are also seeing researchers build bots that monitor programs continuously. Instead of logging into one program at a time and repeating the same recon process, they can track changes across 20 programs, watch for new assets, and identify fresh attack angles as soon as something changes.
This changes the economics of research in three places:
Coverage: automation lets researchers observe more assets and programs for longer.
Timing: continuous monitoring reduces the gap between a change being deployed and a researcher noticing it.
Depth: agents handle repetitive collection and testing, leaving the researcher more time for interpretation and attack-path reasoning.
It also changes what platforms need to support. The future hacker workflow is not a person moving through a static program page. It is a human directing a collection of tools, models and bots that interact with programs and assets continuously.
A single AI vendor can replicate one cognitive style at extraordinary scale. That is useful, but it also creates a monoculture in that the same model tends to follow similar patterns, make similar assumptions, and miss similar things. And, yes, that will evolve as techniques are developed to reduce growing costs.
The researcher community behaves differently. Thousands of people bring different technical histories, cultural contexts, specialist skills, and instincts. They customise models, combine tools in unexpected ways, and approach the same application from different angles. That diversity matters because serious vulnerabilities are often application-specific. They depend on business logic or on chaining several modest weaknesses into meaningful impact.
Automated testing can identify known patterns consistently, run at high frequency, and make a basic level of testing cheaper and more available. But the crowd lifts the ceiling and finds the unusual paths, questions the intended workflow, and notices when two pieces of context do not fit together. When researchers also gain access to better automation, that ceiling rises faster.
This is why scanners, AI pentesting, and bug bounty should not be treated as mutually exclusive choices. Many of our customers already layer them together. Each produces a different kind of signal. The strategic question is how those signals are coordinated and turned into decisions.
CrowdRecon has made this argument much more tangible for us.
A vulnerability report shows the end of a successful research path. It does not show the domains a hacker discovered, the assets they explored, the routes they ruled out, or the areas they returned to. That reconnaissance has always been part of the value created by the community, but most of it was invisible to the customer and to the platform.
CrowdRecon begins to make that work visible. It captures structured context about how trusted hackers explore an attack surface before and between validated reports. That can help security teams understand where human attention is concentrated, where coverage may be thin, and which forgotten or unexpected assets deserve another look. And while it does not guarantee complete discovery or risk reduction, it does validate the premise that the community creates a substantial layer of coverage that outcome-only metrics miss.
Now this is the part I find beautiful. The crowd was already doing the work; researchers were already mapping, filtering, revisiting, and making judgment calls across complex attack surfaces. CrowdRecon does not invent that capability. It helps us see it, structure it, and recognise its value. It also gives us a clearer way to compare complementary sources of coverage.
Automated asset discovery can show what exists.
Scanner data can show where known patterns appear.
Researcher reconnaissance adds context about what skilled adversarial thinkers actually choose to explore and why.
Together, these views can produce a more honest picture of an organisation’s exposure than any single source can provide alone.
More capable researchers and more submissions create a positive problem, but it is still an operational problem. Volume only becomes value when the platform can separate noise from signal, validate findings, and help customers prioritise what matters. That is why our investment is focused across the workflow rather than on one AI feature.
Our draft submission assistant gives researchers earlier feedback, including signals that an issue may be out of scope. Submission throttling helps prevent low-quality, automated reports from flooding triage. Within triage, seven specialised agents can work in parallel on tasks such as duplicate detection, scope checks, domain validation, and reproduction, allowing human experts to focus on impact and judgment.
The principle is the same on both sides: automate repeatable work, expand context, and keep people in control of consequential decisions.
AI is lifting the floor of security testing. It is making known techniques faster, more accessible and easier to run continuously. At the same time, the hacker community is using that same technology to raise the ceiling through creativity and application-specific judgment.
Our latest data shows a community moving from raw acceleration towards more mature orchestration. CrowdRecon shows that its contribution is broader than the vulnerabilities that reach a final report. The crowd also creates coverage, context, and early signals across the attack surface.
The organizations that benefit most will not choose between AI and hackers; they will combine automated breadth with human depth, then invest in the coordination and validation layers that turn both into action. The floor will keep rising, and the remarkable part will be watching the crowd raise the ceiling with it.
