Google's Open-Source Bug Bounty Program Has A Big AI Problem

Bug bounty programs are intended to be mutually beneficial: Major organizations get to plug holes in their security while hackers skilled enough to find vulnerabilities are paid for their effort. However, with the advent of AI, many such programs have been overwhelmed with low-quality reports generated by LLMs rather than human researchers, and often reference fictitious, unexploitable, or hallucinated vulnerabilities.

Rather than helping, this is a hindrance that recently caused Google to join the likes of Curl, Intel, and HackerOne and suspend its paid bug bounty program, called the Open Source Software Vulnerability Reporting Program–OSS VRP for short. These were among the highest-paying bounty programs, meaning ethical hackers now have fewer ways to profit from their talents. Google previously awarded up to $31,337 for the most serious reported vulnerabilities with Intel going as high as $150,000.

OSS VRP isn't gone for good, though. Google says that while the vast majority of automated submissions aren't valid, it has committed to reformatting the program and providing an update in the first quarter of 2027. In the meantime, bounty hunters are encouraged to continue submitting reports to Google's other VRP or Patch Rewards programs, which remain active. Alternatively, you could try claiming Apple's $2 million bounty for discovering exploits that can achieve the same goals as a sophisticated mercenary spyware attack.

It might surprise you to learn that most of what you see online is now AI-generated, including articles, music, social media posts, and comments all "written" by LLMs. Artists rightly criticize much of this avalanche as mediocre at best, but it's a problem for AI companies as well. Researchers warn that AIs overall get worse when learning from AI-generated content, a phenomenon known as the "garbage in, garbage out" rule. Since AI tools like ChatGPT are trained on publicly-available information from around the internet, including from platforms like Reddit, this creates a potentially self-defeating cycle. That's one reason why Google and Anthropic have embarked on large projects to scan physical books: They're untainted, fully human-written content from a time before AI usage was commonplace.

That's not to say there aren't legitimate uses for AI. It's proven extremely effective at detecting some forms of cancer, enhancing accessibility via transcription and live captioning, and creating new medicine for conditions like Parkinson's disease. It even has a place in bug-hunting, having recently identified over 1,000 vulnerabilities in Linux. By comparison, the low-quality reports generated by people who aren't cybersecurity pros are often submitted without extra work to identify a genuine vulnerability from a mistake made by an AI.

AI tools are set to become even more ubiquitous, freely-available, and capable of generating infinite amounts of correct-at-first-glance content. Misuse is likely to spread as well, and humans haven't yet caught up. Companies like Google will likely have to implement even more quality-control systems and dedicate resources to finding the few helpful nuggets among the expanding sea of uselessness.