One bug-bounty triage team says cheap AI reports now act as a denial of service

One bug-bounty triage team says cheap AI-written reports now act like a denial of service on the people who must read them. A speaker on the OWASP Foundation's episode "The AI Surge in Bug Bounty: Insights from Decathlon triage team" made this case from the team's own experience. The transcript does not name the speaker.

What was said

A bug bounty is a scheme where companies pay outside researchers, called hunters, to find security flaws. A triage team reads each report and decides whether it is real, a duplicate or noise.

The speaker said that, according to the platform they cited, almost all hunters now use generative AI. They said hunters run coding assistants to find flaws and submit what comes back.

The team compared two periods. Duplicates rose from seven in 70 days to 37 in 91 days. A new status appeared too, called autoclose. A hunter submits a report without reading it, then closes it right after to avoid losing reputation points. The speaker said all of these were AI-generated.

The speaker also listed signs of AI-written reports. Proof-of-concept sections now hold many more curl requests, which are command-line web requests. Lists are padded to make a report feel urgent and raise its severity. Sentences often say a thing is X, not Y.

The speaker put the cost plainly: "gen AI generated reports without verification and well that's that's a denial of service or or triage service."

The team now closes AI-generated mistakes as spam, with a 10-point penalty instead of three. It also removed technical information exposure and business logic errors from the list that earns rewards. Separately, the platform added a credit system that blocks submissions when a hunter's credits reach zero.

Why it matters

Our reading: an open reporting channel was once limited by the time a person needed to write a report. Generating one now costs little. Checking one still takes a trained person. A report padded with plausible requests is quick to produce, yet it must be tested before anyone can dismiss it.

For managers who run or fund a programme, the budget line is triage time, not only rewards. The speaker said some companies and open-source libraries have stopped their public programmes, and named Intel as one. The speaker gave no reasons for any of them.

The other side

The speaker did not say AI itself is the problem. They said that even with the best model, a hunter who does not understand how it works will likely end up with spam reports. Valid reports were already scarce before the AI surge, by the team's own earlier data.

The evidence is narrow. It comes from one team's tagging of its own reports, and the tips for spotting AI text are informal. The excerpts do not show how many genuine findings AI tools have helped hunters produce. Penalties and credit limits may also hit honest newcomers, who start with few credits.