Google just told bug hunters: stop sending us AI slop
Issue #051 · 4 min read
Google just told bug hunters: stop sending us AI slop
A bounty program froze under fake reports. Plus GitHub's new defenses against the same problem.
The big story
Google just froze its own open source bug bounty program.
Not because it ran out of money. Because of a "significant rise" in AI-generated vulnerability reports that were clogging the pipeline — that's TechCrunch.
Here's the shape of the problem. A bug bounty program exists so outside researchers get paid for finding real holes.
That only works if a human can tell the real reports from the noise. Point an AI model at a codebase, ask it to "find vulnerabilities," and it will confidently describe bugs that don't exist, in code that works fine.
Multiply that by anyone with a free afternoon and a chatbot, and the queue of reports a security team actually needs to read gets buried under reports that just sound right.
This isn't only Google's problem. GitHub has spent the last few days quietly shipping defenses against the exact same failure mode: rate limits on private vulnerability reports so one account can't flood a maintainer, and structured forms that force a reporter to supply a reproducible proof of concept instead of a paragraph of prose — see the rate limits changelog and the structured forms changelog.
Separately, the Rust compiler project SWC has stopped accepting external pull requests altogether, citing AI-generated content — that's the SWC team on Twitter.
For a PM, this is the quiet cost nobody budgets for. If your product has any kind of open intake — bug reports, support tickets, user-generated contributions, code review queues — AI has just made it trivially cheap for someone to flood that intake with plausible-sounding garbage.
The fix isn't "add an AI filter." Google and GitHub are both reaching for the boring tools: rate limits, structured inputs, friction at the point of submission. Worth checking whether your own intake forms could survive the same flood.
What shipped
GitHub Copilot code review now has an API, and "Balanced" is the new default effort level. You can request a review through REST or GraphQL and pick how hard Copilot thinks before it comments — see the changelog. If you've wired Copilot reviews into a CI pipeline, the default just changed under you, so it's worth checking what "Balanced" actually costs in review time before Friday.
GitHub Copilot can now control desktop apps, not just write code. "Computer use" is in public preview for the Copilot CLI and the macOS/Windows app, letting Copilot click around other applications on your behalf — the changelog. This is a meaningfully bigger permission than autocomplete. Anyone piloting it should know exactly what it can touch before turning it on for a team.
GitHub also deprecated a batch of Copilot models this week, across Chat, inline edits, agent mode and completions — the changelog. If your team has prompts or workflows pinned to a specific model name, this is the week those quietly break.
What I'd actually do this week
Audit your own open intake points. Support tickets, bug reports, user submissions — anywhere a stranger can submit free text, assume someone is now doing it with an AI model instead of a brain. Ask whether a structured form or a rate limit would cut the noise, the way GitHub just did.
Check which Copilot models your team's workflows are pinned to. The deprecation list above may already have broken something quietly; better to find it than have a teammate find it for you.
If you're piloting Copilot's new desktop control feature, write down exactly what it's allowed to touch before anyone else on the team turns it on. Public preview is the easiest week to set a boundary, and the hardest week to remember to.
Reply and tell me what's clogging your own intake queue these days.