Decathlon's bug bounty team made AI spam cost something with credits and penalties

A speaker from the Decathlon triage team, on the OWASP Foundation show, described how the team made AI-written spam costly for its senders. Their bug bounty program pays outside researchers, called hunters, to report security flaws. The speaker said weekly invalid reports were divided by two after a credit limit and a narrower rewards list.

What was said

The speaker said the problem is scale. In their words, "almost all hunters use generative AI." The risk is reports sent without checking whether the bug is real. The speaker called that "a denial of service or or triage service," meaning it clogs the people who must read each report.

The team used three measures. The first was a stricter rule on AI spam, in place for about a year. If an invalid report is AI-written, it is closed as spam and costs the hunter 10 points instead of three.

The second was removing some vulnerability types from the list that earns rewards. The speaker named technical information exposure and business logic errors. For the second, the team received dozens and dozens of invalid reports that year.

The third was a credit system from the platform, YesWeHack. A new hunter starts with 10 credits, a veteran has far more, and at zero a hunter can no longer submit. The credits arrived in July, and the list change a few days later.

The speaker compared weekly averages before and after those two changes. The AI-spam penalty was left out of the graph, so its own effect was not measured.

Why it matters

Our reading: when sending costs nothing, the receiver pays. Credits and penalties move part of that bill back to whoever sends the report.

Our reading also suggests why those two classes drew weak reports: they are easy to describe but hard to prove. Removing them from the rewards list is a separate lever from points and credits. So the halving should not be credited to penalties alone.

For managers who run or fund any intake channel, such as support queues or vendor tips, the lesson is about design. This team did not describe relying on one detector. It priced mistakes, capped how many a sender can make, and narrowed what it pays for.

The other side

The speaker named a cost. Accepted reports fell too. The team had not changed scope or budget, so, in the speaker's account, fewer accepted reports was inevitable. The speaker did not say whether any real findings were lost. The speaker also called the graph hard to analyze, so the halving comes from averaging weekly data.

The approach still depends on a judgment that a report was written by AI. The excerpt does not say how reliable that check is, or how it treats a careful hunter who used AI.

The speaker said the team would have moved from a public to a private program if AI-generated reports had kept arriving. That did not happen. It shows the fallback is to shut the door to outsiders.