Anthropic Sends AI Bug Reports With No Human Review

On October 1, Google paused part of its open-source bug bounty, citing a significant rise in automated submissions that were mostly not valid. A week later, Anthropic launched a free service that emails fully model-generated vulnerability reports to open-source maintainers with no human triage at all. Same mechanism, opposite direction, and the difference between them is worth understanding.

What Anthropic launched

Anthropic is unusually direct about the tradeoff it is making. Removing triage enables faster and more frequent scanning, and the company says plainly that it means some reports will be incorrect or invalid. That is a more honest framing than most launches manage.

The number going round needs a caveat

Coverage of the first week describes 77 critical and 803 high-severity findings sent to maintainers, which is where the figure of 880 comes from. Those severities are model-rated and were not confirmed by a person before sending, so 880 verified critical vulnerabilities is not what happened.

The 88% validity figure is real and it comes from somewhere else. Before launch, Anthropic had expert penetration testers review 97 critical and high findings across 48 projects: 85 met its coordinated disclosure bar, 11 described real issues that duplicated known ones, and exactly one was a false positive. That is a strong result on a sample, and it is a sample rather than a measurement of the 880.

Anthropic also notes that maintainers have seldom called a high or critical finding invalid, while some have said the severity ratings can be inflated or that the scanner misunderstood a project's threat model. Reading those two together, the finding is usually a real bug and the label on it is the part to treat as provisional.

Why curl is the test that counts

curl released version 8.23.0 on October 14, fixing CVE-2026-92392, a use-after-free flaw in HTTP/2 handling, with the advisory crediting the discovery to Claude.

That matters because of who the maintainer is. Daniel Stenberg has been among the loudest public critics of low-quality AI-generated vulnerability submissions, and the project closed its bug bounty in January under that pressure. He subsequently credited the scanner with finding one of the more serious curl vulnerabilities reported in recent years.

A sceptic changing position on the evidence is a stronger signal than a benchmark, and it supports the distinction this whole story turns on. The objection was never to machine-written reports. It was to reports that cost more to assess than they were worth.

The bottleneck Anthropic names

The most interesting line in the announcement is about capacity. Over six months of scanning major projects, Anthropic says it discovered more than 29,000 candidate vulnerabilities and was able to manually review and triage roughly 6,000 of them. In its own words, it remains bottlenecked on human capacity to validate findings.

Detection got cheap and validation did not, so the queue grew. The response was to stop validating and ship, which is a defensible choice when the measured validity rate is high and each report carries its own proof.

That shape recurs wherever AI speeds up one step of a process: the constraint moves to the step that still needs a person, and the options narrow to raising validity, attaching evidence so review is cheaper, or accepting a different standard. The general version is in finding the bottleneck after it moves, and the opposite outcome in Google pausing its bug bounty.

What it changes for everyone else

Almost no reader of this maintains open-source software, and almost every reader depends on it, because the libraries being scanned sit inside the products you buy. The visible effect is more frequent patch releases from your vendors over the next year.

Which makes the unglamorous work the work. Know what you actually run, including the components inside purchased software, since most businesses cannot produce that list. Apply updates on a cadence you can defend rather than on the day a headline appears, which is argued in patch management: fast is good, instant is not.

Expect the volume of disclosed vulnerabilities to rise without your exposure having changed, because these bugs were already there and are now being found. A spike in CVE counts is a measurement improving rather than a situation deteriorating, and treating it as the latter leads to panic patching. Sizing that work without a security team is covered in vulnerability management without a security team, and the paid counterpart to this free service in the best vulnerability-hunting AI is now something you buy.

Frequently Asked Questions