Google Suspends Open-Source Bug Bounty Program Amid AI-Generated Spam

Google has suspended its Open Source Software Vulnerability Reporting Program, the company announced this week, joining Intel, Curl, and HackerOne in pulling the plug on what were once among the tech industry's most lucrative bug bounty initiatives. The reason is straightforward and increasingly familiar: the programs have been buried under an avalanche of AI-generated reports that sound plausible but describe vulnerabilities that don't actually exist.

Bug bounty programs operate on a simple premise. A company publishes its code and invites security researchers to find flaws. Those who discover genuine vulnerabilities get paid. It's a transaction that benefits both sides—organizations patch real security holes, and skilled hackers earn money for work they're already doing. Google's OSS VRP paid up to $31,337 for the most serious findings. Intel went higher, offering as much as $150,000. These sums attracted serious researchers who knew how to actually hunt for exploitable weaknesses.

Then came the flood. Large language models can now generate bug reports that read like the real thing—detailed, technical, specific—but reference vulnerabilities that either don't exist, can't be exploited, or are simply hallucinated by the AI. The volume has become overwhelming. Most of the automated submissions, Google says, aren't valid. The company now faces a choice between hiring more people to sort through noise or shutting down the program until it can figure out how to filter the signal from the garbage.

This is a particular flavor of a larger problem that AI companies themselves are beginning to confront. When language models train on publicly available internet data, they absorb everything—articles, social media posts, comments, and increasingly, other AI-generated content. Researchers have identified what they call the "garbage in, garbage out" phenomenon: AI systems trained on AI-generated material tend to degrade over time. Google and Anthropic have responded by investing in projects to scan physical books, which predate the AI era and remain untainted by machine-generated text.

The irony is that AI can be genuinely useful in security work. Machine learning systems have recently identified over 1,000 vulnerabilities in Linux—real, exploitable flaws that human researchers might have missed. AI has also proven effective at detecting certain cancers, improving accessibility through transcription and live captioning, and accelerating drug discovery for conditions like Parkinson's disease. The technology isn't the problem. The problem is that anyone with a laptop can now generate plausible-sounding nonsense at scale, and the people submitting these reports often aren't cybersecurity professionals who would know the difference between a real vulnerability and an AI hallucination.

Google says it will reformat the OSS VRP and provide an update by the first quarter of 2027. In the meantime, researchers can submit reports to Google's other vulnerability programs, which remain open. Apple, for comparison, is still offering up to $2 million for discoveries that could achieve the same level of access as sophisticated mercenary spyware. But the broader pattern is clear: as AI tools become cheaper, more capable, and more widely available, companies will need to build increasingly sophisticated quality-control systems just to separate legitimate security research from the noise. The alternative is to close the door entirely—which is what's happening now.