Apple、AI 生成報告の急増を受けセキュリティ報告数に制限を設ける

Apple now caps how many security reports some researchers can have open at once. And once they hit that cap, they may have to wait 30 days before reporting another potentially dangerous software flaw through the company’s internal security portal.

AI slop floods security pipeline

The company introduced the limitations in June after receiving a flood of AI-assisted reports, many of which turned out to be “AI slop,” meaning they were not genuine vulnerabilities. Although AI can help researchers find possible flaws and compile detailed reports, each submission still has to be reproduced and verified.

Italian cybersecurity company Bynario drew attention to the new rules after it reached the limit. According to reporting from the Financial Times, Bynario used GPT-5.5 through its Atlas platform to find more than 50 possible bugs in Apple’s latest Mac operating system in just three weeks.

One of those findings was a flaw in macOS Screen Sharing that could allow an authenticated VNC user to access protected data and create files with root privileges. Apple assigned it CVE-2026-43760 and fixed it in macOS Tahoe 26.6. Bynario said it couldn’t report the flaw at first because it had already hit Apple’s limit for open investigations. Apple has since reached out to Bynario to review its reports.

A human reviews every security issue reported to Apple, even though the company uses AI to help sort reports when there are too many. If researchers hit the cap, they can ask Apple to raise it. But this shows how quickly AI-assisted research can flood a process that relies on people to check each report.

Apple has deployed the technology internally to find vulnerabilities in its own software. Its latest updates contained about five times as many security fixes as previous release cycles. The same dynamic is playing out across the industry: AI-powered security tools are automating everything from prompt injection testing to agent hardening, compressing work that once took weeks into minutes.

But submission caps don’t distinguish between convincing AI-generated reports and real security flaws found with AI, causing both to compete for the attention of security teams.

Triage can’t keep up

Apple isn’t the only company dealing with this issue. Bug bounty platform HackerOne already uses AI to conduct the first review of incoming vulnerability reports through a service called Hai Triage.

AI can now uncover possible flaws and write them up faster than security teams can determine whether they’re real.

HackerOne’s system reviews a report in minutes and decides if it’s likely valid, invalid, or needs additional checking. It can sort the type of vulnerability, suggest how serious it is, and recommend whether a security analyst should examine it or the report should be closed.

HackerOne says Hai Triage does not make the final call, and all outcomes are reviewed or confirmed by human analysts, who remain responsible for deciding whether a vulnerability is real.

The cost of running these AI-driven workflows is falling fast enough that volume is becoming the default, which only accelerates the triage bottleneck. Using AI to sort these reports could help clear the backlog, but it could also bury a vulnerability if the system mistakes it for “AI slop.”

Researchers with a track record of finding real bugs could be allowed to submit more, while those using AI could be asked to prove the flaw can be reproduced. Apple already asks for affected software versions, steps to reproduce the issue, and a proof of concept or exploit. Without that reference point, the cap could hold up legitimate reports while doing little to stop convincing ones that turn out to be wrong.

Apple’s Security Bounty guidelines warn that researchers who repeatedly submit ineligible reports, including theoretical flaws discovered by AI without proper validation, may have their reports paused for 180 days. Researchers with more than two paused periods may be removed from the program permanently.

Researchers with a track record of finding real bugs could be allowed to submit more, while those using AI could be asked to prove the flaw can be reproduced.

Real bugs already delayed

Apple has been slow to respond to real reports in the past. EasyOptOuts co-founder Tyler Murphy first told the company about a flaw in Apple’s Hide My Email feature in June 2025. The bug could expose the real email addresses the feature was meant to protect. After a year of back-and-forth with Apple, Murphy made the issue public.

Apple said it deployed a patch on July 3, 2026, and fully resolved the vulnerability. However, AppleInsider reproduced the behavior on July 17, two weeks after Apple’s claimed fix. EasyOptOuts has since said the vulnerability is fixed, although addresses exposed before the working patch could remain in email records.

Blunt caps risk real research

That flaw wasn’t part of the current wave of AI-assisted reports, but it shows why limiting submissions alone can be risky. Security teams were already having trouble investigating and fixing real vulnerabilities before AI caused the number of reports to jump.

Apple’s approach to AI extends beyond security. The company recently turned Safari into a platform AI agents can control, which highlights the shift in how it integrates AI into its product infrastructure. That ambition makes the security pipeline problem more urgent: as Apple builds more AI-driven surfaces, it also needs to keep up with the vulnerability reports they attract.

Security teams were already having trouble investigating and fixing real vulnerabilities before AI caused the number of reports to jump.

The post Apple and Bynario agree GPT-5.5 found a real macOS bug. They disagree on the report cap. appeared first on The New Stack.