A legitimate macOS security vulnerability worth as much as $200,000 under Apple’s own bounty rules sat unreviewed for months, according to a report from The Decoder. The reason wasn’t a lack of severity or documentation. It was volume: Apple’s Security Bounty inbox had become saturated with low-quality, AI-generated submissions that made it harder for reviewers to spot the real thing.
The case has become a reference point for a problem security teams across the industry have been warning about with increasing urgency — that generative AI tools have made it trivial to mass-produce vulnerability reports that look credible on the surface but fall apart under scrutiny, and that triage teams are drowning in the noise.
A real flaw, lost in the noise
The researcher behind the discovery reportedly submitted detailed technical documentation of a genuine flaw in macOS, the kind of finding that would normally fall into Apple’s higher payout tiers, which can reach six figures for serious kernel-level or privilege-escalation bugs. Instead of a swift acknowledgment, the report reportedly languished, caught in a backlog padded out by submissions that were vague, inconsistent, or outright fabricated by large language models prompted to describe a security bug.
Security researchers have described this new category of junk submissions as “AI slop” reports — text that mimics the structure of a real vulnerability disclosure, complete with technical-sounding jargon, but that doesn’t hold up when a reviewer tries to reproduce the issue. Because these reports can be generated in seconds and submitted at scale, they can overwhelm the limited number of engineers assigned to triage.
Not an isolated complaint
Apple is far from the only organization facing this. Daniel Stenberg, the maintainer of the widely used curl project, has publicly described a similar flood of AI-generated false vulnerability reports hitting his own bug bounty program, some of which were confident enough in tone to initially fool reviewers before falling apart under basic questioning. Google, Meta, and other major bug bounty operators have quietly acknowledged wrestling with the same dynamic as generative tools become more accessible.
The underlying incentive is straightforward. Bug bounty programs pay out real money for real findings, and platforms like HackerOne and Bugcrowd process thousands of submissions. Even a modest hit rate on AI-drafted reports can be profitable for submitters willing to spam triage queues, especially when payouts for critical flaws run into five or six figures. That math doesn’t require most submissions to succeed — it only requires enough of them to survive an overworked reviewer’s first pass.
Why this matters beyond one bug
The consequences extend past a single delayed patch. Every hour a security engineer spends debunking a fabricated report is an hour not spent validating a real one. When a company’s most sophisticated defense against critical vulnerabilities is a small triage team, junk submissions function as a kind of denial-of-service attack on the review process itself — one that doesn’t need to breach a single system to do damage.
Security professionals have increasingly compared the situation to spam email in its early days: a flood of low-cost, low-effort submissions that forces defenders to build new filtering systems just to find the signal in the noise.
That comparison undercuts one of the traditional strengths of bug bounty programs — the idea that opening a direct channel between outside researchers and internal security teams makes software safer faster. If that channel becomes clogged with noise, the incentive structure meant to reward diligent, unpaid vulnerability hunting starts working against its own purpose.
What companies are doing about it
Some bounty platforms have begun experimenting with automated pre-screening tools designed to flag submissions that show telltale signs of AI generation, such as generic language, missing reproduction steps, or claims that don’t match the actual codebase. Others have discussed tightening reputation systems, requiring stronger proof-of-concept evidence before a report enters the main review queue, or applying financial penalties to accounts that repeatedly submit unverifiable claims.
None of these fixes are simple. Automated filters built to catch AI-generated slop risk also filtering out legitimate reports from non-native English speakers or less experienced researchers who may write less polished submissions. Overly aggressive gatekeeping could discourage the very community bug bounty programs depend on.
The bigger AI-noise problem
The bug bounty flood is one specific example of a broader trend: generative AI lowering the cost of producing convincing-looking content at scale, whether that’s vulnerability reports, spam, or synthetic text designed to pass as human-written. Regulators have started responding to adjacent issues — the European Union’s AI Act, for instance, now requires chatbots to disclose that they are not human, part of a wider effort to keep AI-generated material from blending seamlessly into systems built around human trust.
Apple has not publicly detailed how it plans to prevent similar delays in the future, and it’s unclear whether the researcher behind the $200,000-tier flaw has since received payment or a fix timeline. But the episode adds to a growing body of evidence that AI-generated content isn’t just a problem for search engines and social feeds — it’s beginning to strain the infrastructure that keeps major software platforms secure.
