OSS Scanner sends maintainers 77 critical findings

Anthropic said on X on October 19 that OSS Scanner’s first scans have sent open-source maintainers 77 critical and 803 high findings. The model sets each of those severity ratings itself, and no human reads a report before it goes out. If you maintain a project that infrastructure or user security depends on, enrolling gets you that raw stream as soon as the models produce it, each report with a reproducer and a patch where possible. If your team lacks the people to triage it, Anthropic says its slower, human-verified disclosure process stays in place, especially for projects like yours.
The free, opt-in service was announced in Anthropic’s launch post on October 8. Over six months of scanning major projects with its latest models, Anthropic found more than 29,000 candidate vulnerabilities, but people had reviewed and triaged only about 6,000 of them. More and more maintainers who got the first reports asked for everything else, unverified and with proposed patches, and nearly 5,000 reports have gone out that way. OSS Scanner turns that request into a service: periodic scans by Anthropic’s strongest models, Claude Mythos among them, with output that is fully model-generated.
Each report carries a self-contained reproducer, an explanation of the bug and a candidate patch when one is available. Where possible, the explanation includes a bisection, a search through the project’s history for the change that introduced the bug. Anthropic gives its reason in the post: exploits “can now be developed in minutes,” so projects that find and fix bugs faster are better placed against attackers looking for the same weaknesses.
Who can enroll
Core maintainers of eligible projects enroll by opening a pull request on Anthropic’s GitHub repo, anthropics/oss-scanner, using the standard project template. An extended FAQ has the rest of the guidance. The eligibility criteria are close to those of Google’s OSS-Fuzz, which scans open-source code with fuzzers: a project should have a “critical impact on infrastructure and user security,” and Anthropic decides case by case.
The scanner page sets two terms that matter once reports arrive. Anthropic will not put any form of 90-day coordinated disclosure period on these unvalidated findings, so no public-disclosure deadline comes with them. It also asks maintainers to credit fixes with an ID of the form ANT-2026-ABCD1234.
Two other programmes cover the fixing side. Claude for OSS gives out free Claude Max 20x subscriptions to help remediate vulnerabilities. The Cyber Verification Program gives qualifying security professionals advanced cyber capabilities with fewer blocking classifiers.
How often the raw reports hold up
Before launch, Anthropic had the penetration testers who review its disclosure findings check 97 critical and high findings from an early version of the scanner, spread across 48 projects. Of those, 85 (88%) met the bar for its disclosure process. Of the other 12, 11 were real but duplicated known issues or other findings from the same scan, and one was a false positive. That sample works out to about 11% of the 880 critical and high findings the scanner has reported so far.
Maintainers have seldom told Anthropic that a high or critical finding was invalid. Some have said the severity ratings can be inflated, or that the scanner misunderstood the project’s threat model. Anthropic says it can’t guarantee the scanner will be perfect and will keep refining it based on maintainer feedback.
wolfSSL’s Todd Ouska said that “of the 74 reports we received, all but two were valid, and five became CVEs.” Noah Misch of PostgreSQL said several reports “came with fixes we can use nearly as-is,” and that fast-track access let the project fix the newest issues “before they reached a GA release.” Anton Arapov of OpenSSL Corporation said the reports, raw model output included, “were as good and sometimes better than what we get from people.” During several weeks of pre-launch validation with dozens of projects, Anthropic chained several of the reported vulnerabilities into unauthenticated remote code execution exploits.
Where the queue backs up
As of May 22, Anthropic had disclosed 1,596 findings to open-source maintainers across 281 projects, and only 97 were confirmed patched, the Cloud Security Alliance reported. It concluded that the bottleneck had moved from discovery to remediation, because volunteer maintainers have limited capacity. OSS Scanner takes Anthropic’s human review out of the path, so checking each report now falls to the maintainers who opt in.
curl’s Daniel Stenberg said OSS Scanner helped find “one of the worst curl vulnerabilities reported in the last few years.” Later, on October 20, he announced that from October 26 curl wants Low and Medium vulnerability reports filed as public GitHub issues, with only High and Critical reports kept private on HackerOne. He wrote that “over 90% of all submissions could quickly be assessed to be less than HIGH,” and he described maintainers as “strained, overloaded and brought to the brink.”
Our analysis
What we checked
On October 20 we read curl’s and wolfSSL’s advisory records, counted curl’s full advisory history, cloned the public enrolment repo along with its pull requests, and checked the timestamps on the X posts.
Stenberg’s claim holds up on curl’s own scale: the bug is rated High, only the third High curl has published since 2021. But the timing in the X posts is off. ClaudeDevs presented the curl fix alongside what Anthropic learned in the first week. curl’s advisory, however, says the bug was reported on September 15, 23 days before launch, and @thepunit’s own post says it was found with “an early version of OSS Scanner.”
| Item | What we found |
|---|---|
| CVE-2026-92392 | HTTP/2 upload use-after-free, High on curl’s four-level scale; affects 8.12.0 through 8.22.0; fixed in 8.23.0 on October 14, 29 days after the report, with backports to 8.14.3, 8.16.2 and 8.20.2; curl cut its release cycle from eight weeks to six to ship the fix |
| curl’s record | 2 Critical and 41 High among 239 advisories; 1,099 days since the previous High; kept below Critical because curl deems exploitation “extremely difficult” in the wild |
| curl 8.23.0 | 24 advisories, curl’s largest single-day batch; Anthropic is credited on two, the High one and one Low (CVE-2026-107928), the only curl advisories ever to credit Anthropic or Claude |
| wolfSSL | Five CVEs marked “Found via the Anthropic OSS program,” fixed in 5.9.4: one High, one Medium, three Low |
| Enrolment repo | Zero projects on launch day, 128 at our check; 862 pull requests, 144 merged and 250 open; projects include openssl, openssh, cpython, nss and rails |
Neither curl nor PostgreSQL has a project entry or a pull request in the repo. Our reading of the dates and of @thepunit’s wording is that the curl bug came through the pre-launch validation pipeline. Public records don’t show which channel was used.
Up next
Judge orders ex-OpenAI engineer Alfozan held until trial
On October 4, the FBI says, ex-OpenAI engineer Abdulrahman Alfozan opened a folder named “Saudi Gov Docs” on his iPhone. On Monday a judge ordered him held until trial.
Related stories
- Anthropic cuts deadline for critical bugs from 90 to 30 days
- Claude wrote no-crash proofs for 123 OSS-Fuzz libraries
- Claude Security moves to continuous scans on Enterprise
- Claude found about 4,500 OpenSSH lines reachable pre-auth
- Anthropic’s free OSS Scanner skips human review on purpose
- Testing Catalog: Claude iOS app gets “Saved logins” list
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

