Apple has tightened access to its bug bounty pipeline after a surge in AI-generated vulnerability reports overwhelmed its security review teams.
According to the Financial Times, Apple now limits how many new bug reports a security researcher can keep open at the same time and has added a 30-day cooling-off period between submission batches. Internal researchers can apply in advance for higher limits.
Apple said, "As AI-generated security submissions across the industry have grown, we recently adjusted the number of new reports a researcher can open at one time."
The practical effect is clear: review queues are being flooded by machine-generated reports, making it harder for valid findings to stand out.
Bynario says macOS findings were blocked by the cap
The clearest example in the report came from Bynario, an Italian cybersecurity startup. The company said it used ChatGPT to scan macOS and found more than 50 vulnerabilities in three weeks.
Those results were not all noise. One of them, according to the report, was a privilege-escalation flaw that could give an attacker unrestricted access to a Mac, effectively handing over the whole system.
Bynario said it could not submit the findings because it had already hit its reporting quota. Apple later said it had been in contact with the company and was reviewing its submissions.
The cap creates an obvious trade-off. Apple is trying to block low-quality, AI-assisted volume submissions, but the same limits can also catch researchers who are using AI tools in ways that still produce actionable findings.
curl ran into the same problem
Apple is not the first project to deal with this. Daniel Stenberg, maintainer of the open-source tool curl, shut down curl’s bug bounty program on HackerOne on Feb. 1, 2026 for the same reason: AI-generated reports had become long, confident, and often built on hallucinated claims.
Stenberg said the share of curl submissions confirmed as real vulnerabilities had stayed above 15% for years. In 2025, that figure fell below 5%. By the end of the year, only about 1 out of every 20 to 30 reports was accurate.
He described the workload as less like security triage and more like handling terror-threat reports, draining the time, attention, and "will to live" of his seven-person security team.
curl had offered rewards of $10,000 for critical vulnerabilities and $500 for low-severity issues. Stenberg said that while the payouts were not especially large, they were enough to attract people willing to paste in large language model output and hope something landed.
One month after shutting the program down, curl returned to HackerOne in March. Report quality improved, but the number of submissions kept rising.
Apple is using AI to sort AI reports
Apple’s answer is to deploy its own AI systems for initial classification, while keeping human verification in place.
The company is also using models such as Claude Mythos through its participation in Project Glasswing to proactively look for vulnerabilities in its own products and strengthen defenses against AI-driven attacks.
The picture described in the report is hard to miss: one side uses AI to generate vulnerability submissions, the other uses AI to filter them, and a smaller group of human researchers is left to identify the real bugs in between.
The article also noted that Stenberg had previously called Mythos a "public relations gimmick," while Apple is now using that kind of model as part of its gatekeeping process.