AI Slop Buried a Real macOS Flaw in Apple's Bug Bounty

Alfredo Pesoli, founder of the Italian startup Bynario, was holding a serious macOS flaw, the kind that could hand an attacker full control of a machine, with a black-market value he estimates at $100,000 to $200,000. He couldn't report it. Apple's bug bounty program was so clogged with AI-generated bogus vulnerability reports that the company had capped submissions, and his turn kept not coming.

The twist: Pesoli found the flaw with the help of ChatGPT. The same technology surfaces the real bug and generates the noise that keeps it from being filed.

What clogged the program

According to the Financial Times, Apple's security team was flooded with low-quality reports, largely AI-generated, describing vulnerabilities that don't exist. The company's response was to throttle: a per-researcher submission cap and a mandatory 30-day waiting period for new reports. Proven researchers can request a higher quota; someone new to the program waits at the back of the line no matter what they're holding.

Big noise, small signal

A separate analysis published the same week completes the picture: only 1.3% of the security flaws AI finds turn into a verified exploit. Meanwhile, the time for real flaws to be weaponized has dropped from 120 days to 80. So the vast majority of the report pile is not a real threat, but the minority that is hits faster than before. For a defense that can't separate signal from noise, few combinations are worse: you drown in the pile while the window to act on a real threat shrinks.

Apple uses the same tools

The detail that makes this most interesting: Apple leans on Anthropic and OpenAI models in its own internal security scans, and its latest updates contain five times the usual number of fixes. The issue isn't AI entering security research; it's that nobody yet has a mechanism to filter the AI output coming in from outside. The company closing a record number of flaws internally with these tools is drowning in the reports arriving from outside.

Your inbox is next in line

Reading this as an Apple story sells it short. Few companies run a bug bounty; many receive job applications, support tickets, quote requests, and tender documents, and the flood of AI-generated submissions is coming for those too. What we see in the field: any process that filters incoming content by human effort starts sliding into Apple's position as volume grows. The concrete need for the coming period is filtering layers that prioritize what lands in the inbox, flag the suspect, and surface the real signal. And while building that layer, don't forget the same lesson: if the filter is also AI, measure before you trust it.

Sources: The Decoder, The Decoder (analysis)