- Apple imposed per-researcher caps on open bug reports and a 30-day cooldown period starting in June 2026 after AI-generated submissions flooded its security bounty pipeline [1] - Italian startup Bynario, which had filed over 50 reports in three weeks using GPT-5.5, was blocked from submitting a macOS Screen Sharing vulnerability that CEO Alfredo Pesoli valued at $100,000 to $200,000 on the black market [2] - Apple patched the flaw as CVE-2026-43760 in macOS Tahoe 26 and has since contacted Bynario directly to review its submissions [3] - In the first half of 2026, 1,061 vulnerabilities were traced to AI-assisted discovery, with 14 showing confirmed exploitation — a 1.3% rate [4] - Apple itself credits Anthropic and OpenAI models for identifying real vulnerabilities patched in recent software updates
[3] Apple has imposed caps on the number of security bug reports individual researchers can have open simultaneously, along with a 30-day cooldown period, after a surge of AI-generated submissions overwhelmed the company's review pipeline. The restrictions, implemented in June 2026, are designed to stem what the Financial Times described as a flood of fabricated vulnerability reports dressed up in technical jargon [1].
The new limits produced an immediate unintended consequence. Italian cybersecurity startup Bynario, which had submitted more than 50 potential vulnerabilities in three weeks using an AI-powered platform built on OpenAI's GPT-5.5, was blocked from reporting a genuine privilege-escalation flaw in macOS Screen Sharing. CEO Alfredo Pesoli estimated the vulnerability's black-market value at $100,000 to $200,000 [2].
Apple has since patched the flaw — assigned CVE-2026-43760 — in macOS Tahoe 26 and reached out to Bynario directly. In a statement, the company said it had "adjusted the number of new reports a researcher can have open at once" and noted that researchers can request higher limits to ensure critical findings reach security teams [3].
What Happened #
Apple's Security Bounty program, which has paid out more than $35 million to over 800 researchers since its launch, found itself buried under a new category of submission: reports generated partially or entirely by large language models. Many of these contained hallucinated vulnerabilities — technically plausible on the surface but describing flaws that did not exist [5].
Each submission still requires human review, and the volume exceeded what Apple's security team could process in a timely manner. In response, Apple capped the number of concurrent open reports per researcher and enforced a 30-day cooldown once that ceiling is hit. Researchers who believe they have critical findings can request an exemption to the cap [3].
Bynario's situation illustrated the collateral damage. The firm had used ChatGPT to assist in analyzing macOS code and discovered a flaw in the Screen Sharing feature that would allow an authenticated VNC viewer to access protected data and create files with root-level system privileges — effectively granting full control of the machine. But when Bynario attempted to submit the report, Apple's system rejected it because the firm had already exceeded its allocation [2].
The Paradox of AI in Vulnerability Research #
The Bynario case highlights a paradox at the heart of AI-assisted security research. The same tools that enable skilled researchers to find real vulnerabilities faster also allow less-skilled operators to generate high volumes of superficially credible but ultimately bogus reports. Sorting one from the other requires the same expert human review that the bug bounty model was designed to crowdsource [1].
Rafe Pilling, a security expert at Sophos, told the Financial Times that bug bounty programs have shifted from a discovery role to a validation role, now operating "at machine speed." The challenge is that validation cannot yet be automated with the same confidence [1].
Apple itself uses AI tools from Anthropic and OpenAI internally to identify vulnerabilities, and credited those models for some of the security fixes shipped in recent software updates. The company found five times more security fixes in its latest round of updates, suggesting AI-assisted discovery is genuinely accelerating on both sides of the pipeline [3].
Industry-Wide Problem #
The issue is not unique to Apple. The Financial Times characterized the AI report flood as "a growing crisis for corporate bug bounty programs" across the technology industry. Any company running a public vulnerability disclosure program faces the same triage burden as LLMs lower the barrier to submission [5].
In the first half of 2026, researchers tracked 1,061 vulnerabilities tied to AI-assisted discovery. Of those, 14 showed confirmed exploitation in the wild — a 1.3% rate that underscores both the volume of noise and the real signal embedded within it [4].
Apple raised its maximum bounty payout to $2 million in November 2025 for sophisticated zero-click exploit chains, with individual payouts able to exceed $5 million when bonuses are included. The increased rewards were designed to attract top-tier researchers, but the submission caps risk alienating the same cohort if legitimate findings are delayed [5].
What's Next #
Apple has begun working directly with Bynario to review its outstanding submissions, a stopgap measure that does not scale. The broader question for the industry is whether bug bounty platforms can build automated triage systems sophisticated enough to separate AI hallucinations from genuine zero-days without introducing the kind of delays that leave critical vulnerabilities unpatched [3].
For now, the episode serves as a concrete case study in the second-order effects of generative AI: a technology that simultaneously makes software more secure and the systems designed to report insecurity less functional.
Companies mentioned #
Further sources #
[1] Financial Times, as reported by The Decoder, August 2, 2026 ↗ [2] The Decoder — A real macOS flaw worth $200K went unreported because Apple's bug… ↗
[3] iPhone in Canada — Apple Caps Bug Reports After Surge in Fake AI Security Flaws… ↗ [4] Digital Trends — AI is finding Apple security flaws faster than Apple can sort … ↗
[5] CryptoBriefing — Apple struggles to keep pace with AI-powered bug hunters as vu… ↗ The stories that matter, in one email. Free — unsubscribe anytime.