Apple Caps Bug Bounty Reports Amid AI Hallucinations
On August 2, 2026, Apple limited vulnerability submissions to its bug bounty program after a surge of low-quality reports, including AI-hallucinated bugs, according to the Financial Times as cited by MacRumors and AppleInsider. AppleInsider reports that the restriction includes a 30-day cooling-off period. Cybersecurity firm Bynario said the cap prevented it from submitting a macOS privilege-escalation exploit it discovered using ChatGPT-assisted research.
Apple has limited the number of vulnerability reports security researchers can submit through its bug bounty program after its review system received a high volume of poor-quality submissions, including reports of bugs hallucinated by AI, according to the Financial Times as cited by MacRumors and AppleInsider.
AppleInsider reports that the controls include a 30-day cooling-off period for submissions through Apple's security portal. The restriction is intended to manage a review backlog, but it also affects researchers submitting potentially valid reports.
A valid-report bottleneck
Italian cybersecurity firm Bynario told the Financial Times that it used ChatGPT-assisted research to identify more than 50 macOS bugs in three weeks. MacRumors reports that Bynario had submitted eight reports in 2025 and five in 2026 before reaching Apple's reporting limit.
Among Bynario's findings was a privilege-escalation exploit chain that could provide unrestricted access to a Mac, according to MacRumors. The firm was unable to submit that report after reaching the cap. Apple told the Financial Times that it was in contact with Bynario and reviewing its submissions, AppleInsider reports.
The reports distinguish between AI-assisted discovery by established researchers and unverified submissions generated or amplified by less experienced users. AppleInsider characterizes the latter as "AI slop": reports that appear plausible but do not describe reproducible vulnerabilities.
Disclosure systems face an automation problem
Across coordinated vulnerability-disclosure programs, automated code analysis can increase both the discovery rate and the triage burden. The practical constraint is not only whether a model can propose a flaw, but whether the reporter can provide a reproducible proof of concept, affected versions, exploitability evidence, and a clear security impact.
For security teams, the Apple case places greater value on submission-quality controls that separate high-confidence findings from speculative model output. Common mechanisms in comparable programs include researcher reputation signals, reproducibility requirements, duplicate detection, rate limits, and escalation channels for vulnerabilities with credible evidence of severe impact.
Apple's reported cap also exposes a tradeoff: broad limits can reduce reviewer overload, while potentially delaying a valid report from a researcher who has already reached the threshold. The sources do not specify the cap's numerical limit or describe how Apple prioritizes submissions within the restricted workflow.
Key Points
- 1Apple introduced report limits and a 30-day cooldown after AI-generated and low-quality vulnerability submissions increased review workload.
- 2Bynario reported that the cap blocked submission of a macOS privilege-escalation exploit, illustrating how rate limits can affect valid findings.
- 3Across disclosure programs, AI-assisted discovery makes reproducibility, proof-of-concept quality, and triage automation increasingly important operational controls.
Scoring Rationale
The report concerns a major platform vendor's vulnerability-disclosure workflow and the operational security consequences of AI-generated submissions. It is relevant to security researchers and teams building automated vulnerability discovery pipelines, though it is not a new model, exploit disclosure, or broadly deployed technical release.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
