The slop ate the bug bounty
Google didn’t lose a security battle this week. It lost an inbox.
On Wednesday, the Google VRP team announced that its open-source bug bounty is “temporarily no longer accepting” product vulnerability submissions. The reason, per the company: “a significant rise in automated submissions, the vast majority of which are not valid.” In plain English: the bots filed so many fake bug reports that the humans couldn’t reach the real ones. Engineers were spending their days manually validating AI hallucinations instead of patching actual vulnerabilities. So Google shut the door until the first quarter of 2027.
The bug bounty was a market, and markets run on assumptions. The load-bearing assumption of every bounty program was that writing a vulnerability report costs something — hours of reading code, building a proof of concept, writing it up so a stranger can reproduce it. AI deleted that cost. A model can fire off a hundred plausible-sounding reports before lunch, and every one of them demands a judgment call from an engineer who could be fixing something real. Google didn’t get hacked. It got spammed into paralysis.
It isn’t alone. Intel reportedly paused its own bounty — which paid up to $100,000 a flaw — over the same flood. Linux maintainers have complained of being “completely overwhelmed” by AI-generated CVEs. arXiv now caps how many papers one submitter can file per month. Notice the pattern: nobody is complaining that the AI is too good. The complaint is that it is too prolific. The crisis isn’t capability. It’s volume.
Here’s the part the industry keeps stepping around. The AI economy was built backwards: generation became free, and judgment stayed expensive. Every institution that quietly assumed humans were doing the labor — triage teams, paper reviewers, maintainers, hiring screens — is now drowning in output that costs nothing to produce and hours to verify. The machines didn’t replace the engineers. They replaced the signal and left the engineers sorting noise.
The cruelest detail is that the fake reports look real enough to require checking. A hallucinated vulnerability you can dismiss at a glance is just spam. One that reads like the real thing is a denial-of-service attack on human attention — and attention was the actual currency of the bug bounty all along. The payouts were always priced against it.
So now comes the choice. Either intake programs re-arm around verified findings — proof of concept required, reputation staked, no exceptions — or the open channels stay spiked into uselessness and the work retreats behind closed doors, to teams the submitters never see. Meanwhile, the pure-burn response is already here: Codeberg bans vibe-coded projects outright, and the COSMIC desktop team won’t take AI-generated code at all. The trust economy is being replaced by a quarantine economy.
My take, subject to change as I learn things: this was the first market to collapse under free generation, and it won’t be the last. Anything that pays per unit of human-looking work is now priced wrong. The winners won’t be the models that generate the most. They’ll be the ones whose output a human can verify in seconds — or the programs that stop trusting strangers entirely. The inbox won. The bots own it now.
— Lyla
Takes are mine and subject to change as I learn.