Anthropic's Claude Surfaces 493 Issues in 25 OSS Projects

Anthropic's Frontier Red Team pointed Claude at 25 open-source projects totaling over five million lines of code, and the model flagged 493 issues, by the company's own count.

The writeup, authored by Chris Rohlf, defines the exercise as "the process of using AI models to map the code that is reachable, then removing or restricting access to anything than a given deployment doesn't need to expose." Claude catalogued more than 17,000 configuration options across the projects and classified 3,842 as security-relevant. In 162 cases, pairs of options weakened security when combined; in 94 of those, Rohlf writes, "the combination silently turned off a security control."

The OpenSSH case study carries the sharpest number. Claude "traced the input paths of the SSH daemon in the 10.4 release and found roughly 4,500 lines of new and experimental code for ML-DSA-44," a post-quantum signature algorithm. Those lines, the post states, are "about 16% of OpenSSH's own pre-authentication code in a default sshd." OpenSSH 10.6 incorporated the resulting fixes, which the post says "removed roughly a fifth of the OpenSSH code that an unauthenticated client previously could have reached."

Anthropic is explicit about the ceiling of the method. "It is perhaps impossible to measure how many unknown vulnerabilities are in any given codebase," the post notes, adding that these techniques "don't remove the unknown vulnerabilities in the code, but they do make the vulnerabilities harder to reach." The write-up publishes the aggregate issue counts and the OpenSSH figures, but no per-project severity breakdown, no CVE assignments, and no accepted-upstream tally beyond the one SSH release.

The post sits in a dense run of Anthropic research output over the past weeks, with Claude cast as the instrument rather than the product.