Claude cuts OpenSSH pre-authentication attack surface by about a fifth; 493 issues flagged across 25 projects
Sourcing note: this report rests on a single reporting chain. The findings come from Anthropic's own Frontier Red Team research post of October 20, 2026, and the coverage found, including AI Weekly, reports that post. The issue counts are Anthropic's own, the work has not been peer reviewed or independently replicated, and Anthropic states that none of the 493 issues was immediately exploitable on its own. The one independently visible result is upstream: OpenSSH shipped the resulting fixes in its 10.6 release. This item will be updated when independent results appear.
Anthropic's Frontier Red Team ran Claude over 25 open source projects totalling more than five million lines of code, cataloguing more than 17,000 configuration options, 3,842 of them security relevant. Claude flagged 493 attack surface issues, including 162 pairs of configuration options that weakened security when combined, 94 of which silently turned a security control off, and 93 unnecessary pre-authentication surfaces in the default configuration of the eleven login gated projects. In OpenSSH 10.4, Claude traced roughly 4,500 lines of experimental ML-DSA-44 post-quantum signature code that an unauthenticated client could reach even when the algorithm was disabled, about 16 percent of the default pre-authentication surface. Fixes developed with the OpenSSH maintainers and shipped in OpenSSH 10.6 removed roughly a fifth of the code an unauthenticated client could previously reach. Anthropic is packaging the method as a reusable Claude skill for its OSS Scanner.