Anthropic Adds Opt-In Security Scanning for Critical Open-Source Projects
Oct 23, 2026
AI company Anthropic has launched OSS Scanner, a free service that helps find security problems in open-source software projects. Project maintainers, the people who look after the software, can choose whether to join. Those who sign up receive regular security reports made by Anthropic's most powerful AI models, including Claude Mythos. The reports are sent without a person checking them first. Maintainers can check the findings themselves and decide which suggested fixes they want to use.
In an announcement on October 8, Anthropic's Frontier Red Team explained that checking AI findings by hand has become a problem because it takes so much time. Over six months, its AI models found more than 29,000 possible security weaknesses. But human researchers were able to review and sort through only about 6,000 of them. Some maintainers had already asked to see the results sooner. Anthropic had sent them nearly 5,000 reports that had not yet been checked, along with suggested fixes.
Each scanner report includes a small, complete example that shows how to make the suspected security problem happen. It also explains what the problem is. When possible, the scanner uses a method called bisection to find the code change that first caused the bug. It also suggests a possible fix. This builds on Anthropic's earlier work through Project Glasswing. InfoQ reported on that project in April, when Anthropic limited access to Mythos Preview to organizations taking part in the project.
PostgreSQL maintainer Noah Misch said that getting early access to the reports helped the project fix problems before a release became available to everyone:
"Several reports came with fixes we can use nearly as-is"
Noah Misch, PostgreSQL
Anthropic tested an early version of the scanner by asking security testers to check 97 findings that were rated critical or high severity. These findings came from 48 projects. Of the 97 findings, 85 met Anthropic's rules for coordinated vulnerability disclosure, a process for telling software makers about security problems so they can fix them. Another 11 findings were real problems, but they repeated issues that were already known or had been found by other scans. One finding was not valid. These results only show how well the scanner worked on this particular group of findings. They do not prove how accurate every future report will be.
Anthropic has also admitted that some maintainers have received reports that made security problems seem more serious than they really were. Some reports also misunderstood how attackers might target the projects. Even so, Anton Arapov of OpenSSL Corporation said the reports were "as good and sometimes better than what we get from people". He especially liked reports that included a working example of an attack, because those made it easier to check whether the problem was real.
Todd Ouska of wolfSSL said the suggested fixes worked well with the project's normal process for checking and fixing security problems:
"of the 74 reports we received, all but two were valid, and five became CVEs."
Todd Ouska, wolfSSL
There is also an example confirmed by a maintainer outside Anthropic's announcement. Daniel Stenberg, who maintains curl, said Luna Tong at Anthropic found a security problem using OSS Scanner. The problem involved a use-after-free bug during HTTP/2 uploads. This kind of bug happens when software tries to use memory after it has already been released. The issue, called CVE-2026-92392, was fixed in curl 8.23.0. Under certain conditions, it could let an attacker run code remotely on another computer. Stenberg rated the problem as high severity rather than critical because it would be difficult for an attacker to use it successfully. He said the more likely result would be a program crash.
Anthropic says Google's OSS-Fuzz helped inspire OSS Scanner. However, OSS Scanner uses AI language models to examine code. Another similar tool is Google DeepMind's CodeMender, which InfoQ wrote about in October 2025. In the announcement introducing CodeMender, researchers Raluca Ada Popa and Four Flynn explained how the system automatically checks suggested fixes by analyzing programs and running tests. CodeMender still required a person to review fixes before sending them to the original software projects. OSS Scanner is different because it sends findings directly to maintainers without human review.
People in the software community have also discussed the work that comes after a security problem is found. Varun Badhwar wrote on LinkedIn: "An upstream patch doesn't help much if you're stuck three major versions behind." Badhwar is the CEO and co-founder of Endor Labs, a company that makes application security tools. He argued that companies need to check whether attackers can actually reach the vulnerable code. They also need to know whether they can install a fix without causing their applications to stop working.
Dimitrios Thomas wrote on LinkedIn, sharing his personal views, that people should judge these tools by "validation accuracy, maintainer workload and time from discovery to a deployed fix". His suggested ways of measuring success look at more than just how many security problems a tool reports.
The OSS Scanner FAQ explains that AI agents check bugs again, suggest fixes, and study the underlying causes before reports are sent by email. These checks are done automatically by AI, not by people. Later scans look for new problems added to the software and older problems that previous scans missed. How often a project is scanned depends on several things, including how widely the project is used and how much work the scanner service is handling.
Core maintainers can sign up by submitting a GitHub pull request. The request must include settings for their project and a Dockerfile that explains how to prepare the software and the other software it needs to run. After the initial setup, scanning runs without an internet connection. Maintainers can also add an optional file called a threat model. This file explains what the scanner should look at and how serious different security problems might be. Anthropic decides which projects can join one at a time. It gives priority to well-established projects that are important for computer infrastructure and user security, especially projects that already have a way to handle confirmed reports of serious security problems.
Maintainers can use the threat-model file to explain which information coming into their software should not be trusted, which parts of the software should be left out of scans, and how they decide whether a security problem is serious. The file can also describe how maintainers want reports to be written and what they expect from suggested fixes. Anthropic checks every sign-up request by hand to make sure the person applying is a core maintainer. It also checks that reports will go to someone officially allowed to handle security issues for the project.
Projects can stop receiving reports for a while or leave the service completely. Reports that have not been checked do not automatically start a 90-day deadline for publicly revealing a security problem. However, if someone later checks a report by hand and confirms that the problem is real, it can enter a separate coordinated disclosure process. In that case, the 90-day period begins when maintainers are told that the problem has been confirmed. Anthropic will also continue sending security reports that have been reviewed by people, especially to projects that do not have enough time or resources to check large numbers of unverified findings themselves.