Every few weeks, someone predicts AI is about to make human security researchers redundant. The logic is tidy: AI can map an attack surface, flag known vulnerabilities and automate the reconnaissance that used to fill a researcher’s week. If a machine finds bugs faster and cheaper than a person, why keep paying the person?
It sits in the half of the work the tidy logic ignores: the vulnerabilities that carry real business risk are rarely single, obvious flaws, and finding them takes adversarial creativity and many angles of attack, not finite scripted passes.
What changes in bug bounty is the toolkit, and frontier AI models are the latest chapter. They’re genuinely powerful against known and emerging vulnerability classes, and organizations are right to use them: machines work faster than any human researcher, and every exposure a defender finds and fixes first is one an attacker never exploits. Attackers run the same tools; this isn’t optional.
AI Has Automated The Floor
Basic bug hunting is becoming commoditized. AI identifies and increasingly drafts fixes for known patterns cheaply and at scale. It’s strong at variant analysis, too: once one flaw surfaces, AI can sweep a codebase for every other instance in minutes – work that used to cost a researcher days. For most organizations, that’s good news.
But the effects aren’t all good. The first is noise: when anyone can point an AI tool at a target, low-quality submissions climb fast. Those are plausible-looking findings filed without confirming they’re exploitable or showing how to reproduce them. Programs spend real effort separating genuine findings from guesswork, and those that can’t are drowning in it.
The second problem is more dangerous because it hides. Sound security depends on independence: whoever checks the work isn’t whoever did it. AI doesn’t retire that principle; it makes it harder to satisfy. When the same model writes the code, reviews it and clears it, the check isn’t independent; a model grading its own work tends to agree with itself, and the bugs it misses stay missed.
The deeper danger is correlation: when everyone leans on the same handful of models, they inherit the same blind spots at once. A blind spot shared across an industry is exactly what capable attackers hunt for. The point isn’t that machines can’t review; it’s that the review has to be genuinely independent: a different model, a different method or a different mind.
Finding Bugs Was Never the Hardest Part
Speed of discovery only helps if the fix keeps pace, and it usually doesn’t. Now that finding is cheap, the constraint has moved downstream. A report isn’t risk reduced; it’s risk identified. Turning one into the other is where programs stall today.
The pattern shows up across the industry, and in our own platform data. Over the past year teams got far faster at the repair itself, cutting time-to-fix on critical vulnerabilities by roughly 70 percent. Yet finding accelerated faster still, and the backlog of unresolved critical bugs grew. The bottleneck isn’t the speed of the fix. It’s everything upstream: confirming the flaw is real, judging what it threatens, routing it, prioritizing it. Finding vulnerabilities got faster. Fixing fell behind.
AI helps here, but it can’t close the gap alone. What to fix first relies on human-owned context about what the business can least afford to lose. AI can draft the patch but not certify it, for the same reason it can’t safely review its own code.
Human Work is Moving Up, Not Away
Today’s most advanced cyber-capable models are formidable: they find vulnerabilities and assemble exploit chains on their own, and reach further with every release. But what they reach for still traces the shape of what they were trained on. Run one a hundred times and you get variations of one kind.
Thousands of independent researchers give you a different kind entirely: the business logic flaw no scanner understands, an intent the training never encoded. These are starting points no model shares. And the ground keeps expanding, because the attack surface grows with every AI system deployed.
The money tracks this. If AI were pushing researchers aside, payouts would be falling. They aren’t: in the first half of this year, researchers earned more than $47 million through the H1 Bounty Platform, up more than 25 percent year on year. That growth isn’t evenly shared; as automation crowds the routine end, easy findings pay less, and those pulling ahead do work the models cannot. Demand for human ingenuity is rising, not thinning out.
That puts an obligation on the platforms, not the researchers. When anyone can generate a plausible-looking report, a triager’s attention is the scarce resource, and a good finding shouldn’t wait behind a hundred automated guesses. So this year we changed our triage logic to prioritize researchers with a strong signal: a track record of valid, reproducible, non-duplicate findings. The tradeoff is that lower-signal submissions wait longer.
That’s the right call only if the ladder is fair: signal earned on what researchers submit, not seniority or volume, so a sharp newcomer climbs fast and a big name coasting doesn’t hold a place they no longer earn. It also has to be something a researcher can see, rebuild and not lose to an unfair close. Get it right and the value automation frees reaches those doing what machines cannot; get it wrong and we’ve built a faster gate. We intend to get it right.
So, this was never AI or humans. The strongest security pairs the reach of frontier models with the ingenuity of a global researcher community. AI raises the floor. Human ingenuity raises the ceiling.
