Picture the queue. A language model files a few hundred bug reports at your security portal in a week. Most are hallucinations. A few are real. What do you do?
Apple's answer, per The New Stack, is a rate limit. The company has started capping the number of security reports some researchers can have open at any one time in its disclosure portal. Hit the cap and the wait before submitting the next one can stretch to 30 days. The article says the limit was introduced in June, after a flood of AI-assisted reports arrived that Apple concluded were mostly not genuine vulnerabilities. The piece's word for the traffic is "AI slop."
The framing device in the article is a live disagreement. Apple and a researcher the piece names Bynario both agree that GPT-5.5 flagged a real macOS bug. They do not agree on the cap. That is the whole story in one sentence.
The lever Apple actually pulled
Notice what the change is, and what it is not. It is not a stricter proof-of-concept requirement. It is not a machine-readable submission format. It is not a per-researcher quality score that penalises the same account twice for the same false positive. It is a queue-length limit with a cooldown. A rate limit, applied to humans (even when the human is holding a language model).
Rate limits solve one problem cleanly: they cap the wall-clock cost of servicing bad input. They solve nothing about the quality of that input. If nine of every ten reports are noise before the cap and nine of every ten are noise after it, the security triage team's signal-to-noise ratio has not moved. What has moved is the throughput.
That is a defensible choice. Apple has finite triage capacity, and every hour spent chasing a hallucinated buffer overflow is an hour not spent on the bug that ships in the next OS build. But do not confuse throughput management with quality management. They are different levers, and pulling one is not pulling the other.
Why a CI/CD person should care
If you run any pipeline that ingests AI-generated artifacts and hands them to a reviewer, you are Apple's triage team on a smaller scale. Autofix comments on your PRs. An agent that opens SAST tickets against the default branch. A coding agent that files a dozen refactor PRs before lunch. Each is a queue whose producer got dramatically cheaper while the reviewer did not.
The two responses on the table are the same two Apple weighed:
- Rate limit the producer. Cap concurrent agent PRs per repo. Throttle autofix suggestions per file per week. Require a human sponsor before an agent opens a new session. Cheap to build. Buys reviewer capacity back. Does not improve the average submission.
- Quality-gate the producer. Require every AI-generated finding to carry a reproducing test, a proof-of-exploit, or an attestation from a validator agent. Expensive to build. Targets the actual disease.
Most teams reach for (1) first because it is a one-line policy. Fair enough. Just book time for (2), because the producers get faster every quarter.
The concession the piece deserves
To Apple's credit, capping the queue is honest. It says out loud that a security portal has a throughput budget and that AI-assisted researchers can blow it. Plenty of vendors would have quietly extended their SLA and blamed staffing. Publishing the name "AI slop" is also a small act of clarity that other bug-bounty programs will feel free to echo.
The Apple-versus-Bynario disagreement is real: a legitimate finding, from an AI-assisted researcher, sits inside the same population the cap is designed to slow down. There is no clean way to sort those two apart at intake. That is the shape of the problem, and any team building the next SAST-plus-agent workflow will meet it too.
What to copy and what to leave
Copy the honesty. Write down, in-repo, how your team handles AI-generated PRs, findings and reports. Reviewers cannot triage what they cannot count.
Do not copy the rate limit as a substitute for a rubric. Ship a rubric first. Then, if the queue still overflows, add the rate limit on top of it.
The next AI-assisted researcher you hear about will file more reports before your morning stand-up than your last team could in a quarter. Plan for that queue. Or wait, 30 days at a time, for it to plan for you.
Top comments (0)