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

Anthropic launched OSS Scanner, a free opt-in vulnerability identification service built to inspect critical open-source codebases using frontier models such as Claude Mythos. Announced on October 8 by Anthropic's Frontier Red Team and detailed in InfoQ's report, the system delivers machine-generated security reports straight to registered maintainers without manual human filtering. Over a six-month evaluation, Anthropic models pinpointed more than 29,000 potential security issues, but human researchers were only able to triage around 6,000. Maintainers requesting early access received nearly 5,000 direct unverified dossiers alongside suggested remediations, prompting the team to transition from internal manual bottlenecks to direct project-level automated delivery.

This report examines how Anthropic structured the opt-in scanning architecture, the verification metrics reported across major systems like PostgreSQL, OpenSSL, and curl, and how core maintainers register their repositories. The coverage is tailored for software maintainers, devops engineers, and open-source infrastructure teams evaluating automated model auditing against their codebases. Maintainers retain complete control over validating incoming findings and deciding whether to merge candidate patches into upstream trees.

What shipped in Anthropic Opt-In Security Scanning

The core delivery mechanism in OSS Scanner is an automated notification pipeline sending structured vulnerability packages directly to enrolled core maintainers. Each delivered report contains an explanation of the suspected vulnerability, a self-contained reproduction harness, the specific commit that introduced the fault isolated via automated bisection, and an initial patch candidate. The workflow builds directly upon Project Glasswing, an earlier initiative where access to Mythos Preview was limited to selected partner organisations.

Enrolment operates through a GitHub pull request submitted by repository owners, supplying configuration settings and an isolated Dockerfile specifying how to compile the target program and bundle dependencies. Once setup finishes, scanning runs inside network-isolated environments lacking external internet access. Repository teams may optionally supply a custom threat-model document specifying in-scope input vectors, components to exclude from scanning, preferred notification structures, and criteria for rating vulnerability severity.

What improved in Anthropic Opt-In Security Scanning

Anthropic overhauled its delivery capacity by automating the triage step previously stalled by human researcher capacity limits. In earlier trials, penetration testers examined a sample of 97 high and critical severity findings distributed across 48 target projects to evaluate discovery precision against strict coordinated vulnerability disclosure criteria. While the automation previously struggled with inflated severity designations and mismatched external threat assumptions, automated agent pipelines now execute internal root-cause analysis, bug cross-checking, and patch drafting before emailing output to project maintainers.

The pilot demonstrated significant utility across established infrastructure projects. At wolfSSL, co-founder Todd Ouska reported that out of 74 automated scan packages delivered, 72 proved valid and 5 resulted in published CVE identifiers. In curl, researcher Luna Tong used the scanner to identify CVE-2026-92392, an HTTP/2 upload use-after-free fault patched in curl 8.23.0 that carried remote code execution implications, which Daniel Stenberg evaluated as high severity rather than critical due to exploitation hurdles.

What you gain from Anthropic Opt-In Security Scanning

Open-source engineering leads receive fully packaged reproduction environments that remove the traditional burden of synthesizing test harnesses from abstract bug reports. PostgreSQL committer Noah Misch noted that early delivery windows allowed maintainers to incorporate submitted code fixes almost directly into pre-release trees ahead of scheduled general-availability milestones. By providing exact bisection commits, the service cuts down investigative overhead across multi-decade code trees.

OpenSSL Corporation maintainer Anton Arapov noted that raw automated packages proved as reliable and occasionally superior to typical external human submissions, especially when supplied with functional proof-of-concept exploits. Furthermore, findings arriving through the unverified channel do not impose rigid 90-day coordinated vulnerability disclosure deadlines. Maintainers evaluate candidate issues according to their own release velocity, and the 90-day disclosure clock only activates if Anthropic later conducts a separate manual review and officially notifies the project team.

How to get Anthropic Opt-In Security Scanning

Project maintainers join by submitting a repository PR to Anthropic containing build instructions and verification metadata. Anthropic manually reviews each PR to verify that applicants hold core committer status and that security disclosures route to authenticated private contacts. Maintainers working with CLI configurations, test environments, or experimental auditing run standard package management updates:

npm install -g @anthropic-ai/claude-code
claude update

When switching scanning or interaction engines to frontier preview instances, maintainers select the appropriate model in session or launch pinned sessions directly from terminal prompts:

/model claude-fable-5-1
/model fable

claude --model fable
claude config set model fable

What to watch after Anthropic Opt-In Security Scanning

While rapid generation expands vulnerability visibility, community security leads emphasize that raw detection does not resolve deployment constraints. Application security vendor Endor Labs co-founder Varun Badhwar noted that upstream code modifications provide minimal relief to enterprises maintaining applications pinned three major versions behind, where reachable execution pathways must be confirmed before merging changes. Software engineer Dimitrios Thomas observed that long-term utility will hinge on maintainer review friction, discovery-to-patch timeframes, and true verification accuracy rather than simple volume metrics.

Maintainers must also watch for periodic recurring scans, which Anthropic schedules dynamically based on model availability, cluster workload, and code release cadence to catch newly committed bugs or subtle historical defects. For small development teams unable to process high volumes of raw findings, Anthropic maintains its conventional human-vetted disclosure pathway. Enrolled projects retain the unilateral right to update their custom threat models, pause incoming dispatches, or exit the automated program altogether.