AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Anthropic Adds Opt-In Security Scanning for Critical Open-Source Projects

Collected Oct 23, 2026

Anthropic has introduced OSS Scanner, a free, opt-in vulnerability-hunting service for open-source projects, announced on 8th October by the company's Frontier Red Team. Enrolled maintainers receive periodic security reports produced by Anthropic's strongest models, including one named Claude Mythos. The reports go out without human review; maintainers themselves decide whether to validate findings and which proposed fixes to accept.

Anthropic frames the launch around a bottleneck in validation rather than in discovery. Over six months, the company says its models surfaced more than 29,000 candidate vulnerabilities, but researchers were able to manually review and triage only about 6,000 of them. Maintainers who asked for earlier access had already received nearly 5,000 unverified reports along with proposed patches, a preview of the model the service now formalises.

Each report is meant to stand on its own. It includes a self-contained reproducer and an explanation of the suspected flaw. Where the tooling can, it identifies the change that introduced the bug through bisection, and it supplies a candidate patch. The work builds on Project Glasswing, which InfoQ covered in April when Anthropic limited Mythos Preview access to participating organisations.

The early results are mixed but encouraging, according to the maintainers quoted. PostgreSQL's Noah Misch said fast-track access helped the project deal with issues ahead of a general-availability release, noting that several reports came with fixes the team could use almost unchanged. wolfSSL's Todd Ouska said candidate patches fit the project's existing verification and remediation process, and that of 74 reports received, all but two were valid, with five becoming CVEs. OpenSSL Corporation's Anton Arapov called the raw reports as good as, and sometimes better than, what the project gets from people, especially when a working exploit makes verification straightforward. Anthropic also acknowledges that some maintainers have seen inflated severity ratings and misunderstandings of projects' threat models.

Anthropic tested an early scanner version by having penetration testers assess 97 critical and high-severity findings across 48 projects. Eighty-five met its coordinated vulnerability disclosure criteria, eleven were real but duplicated known issues or other scan findings, and one was invalid. As the company itself frames it, those figures describe a selected sample and do not establish an accuracy rate for every report the service will produce.

A maintainer-confirmed example sits outside the announcement: curl's Daniel Stenberg attributed an HTTP/2 upload use-after-free to Luna Tong at Anthropic using OSS Scanner. Fixed in curl 8.23.0, CVE-2026-92392 could allow remote code execution under specific conditions, though Stenberg rated it high rather than critical because exploitation would be difficult and a crash was the likelier outcome.

The scanner's internals rely on agents that double-check bugs, propose patches and analyse root causes before reports are emailed. That automated checking is distinct from human review. Later scans look for newly introduced problems and issues earlier passes missed, with frequency depending on factors such as how widely a project is used and the service's workload. Scanning runs without internet access after setup.

Onboarding is deliberately manual. Core maintainers enrol through a GitHub pull request containing project configuration and a Dockerfile that prepares the build and its dependencies. An optional threat-model file can document scope and severity expectations, flag untrusted inputs, exclude components, state how the project judges severity, and specify preferred report formats and patch expectations. Anthropic reviews requests by hand to confirm the applicant is a core maintainer and that findings will reach an authorised security contact. Eligibility is decided case by case, with priority for established projects important to infrastructure and user security that can already handle verified high-severity reports.

Projects can pause reports or withdraw. Unvalidated findings carry no automatic 90-day disclosure deadline; if a report is later validated manually, it can enter the separate coordinated disclosure process, with that 90-day period starting when maintainers are told of the validation. Anthropic continues to offer human-reviewed disclosures, particularly for projects without the resources to triage raw findings.

In context, the design differs from two well-known neighbours. Google's OSS-Fuzz, which Anthropic cites as an inspiration, uses fuzzing rather than language models to find crashes. Google DeepMind's CodeMender, covered by InfoQ in October 2025, described automated patch validation through program analysis and testing and kept human review before upstream submission. OSS Scanner instead delivers unreviewed findings directly, which puts the triage burden on maintainers but removes a review queue that Anthropic says it cannot clear at the rate its models generate candidates.

Community commentary points at the work after discovery. Varun Badhwar, CEO and co-founder of Endor Labs, argued that an upstream patch does little for organisations stuck several major versions behind, and that enterprises need to establish whether vulnerable code is reachable and whether a fix can be adopted without breaking applications. Dimitrios Thomas, writing in a personal capacity, suggested evaluations track validation accuracy, maintainer workload and time from discovery to a deployed fix rather than raw report counts.

A likely trade-off is that the same automation that lifts the volume of findings also raises the cost of false positives for volunteer maintainers, who may lack the time to reproduce and dismiss weak reports. Enrolment screening and the threat-model file look aimed at that risk, letting projects steer the scanner away from components and inputs that generate noise. The penalty structure is also softer than a conventional disclosure programme: unreviewed reports cannot start a disclosure clock on their own.

Why it matters: Critical open-source projects are frequently maintained by small teams with little security capacity, and a free model-driven scanner with reproducers and candidate patches could shorten the path from a latent bug to a shipped fix. For developers and product teams, the practical effect is that more unverified security reports will circulate among upstream projects, so consuming and triaging those reports quickly, and being able to adopt fixes rather than sit on old versions, becomes more important. Whether the approach scales depends on maintainer capacity to validate what the models produce, which is an inference from the stated validation bottleneck rather than a result Anthropic reports.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

AI company Anthropic has launched OSS Scanner, a free, opt-in service to find vulnerabilities in open-source projects. Enrolled maintainers receive periodic security reports generated by the company's strongest models, including Claude Mythos. The reports are sent without human review, and project maintainers can choose to validate their findings and decide which proposed fixes to accept. By Matt Saunders