Claude attack surface reduction: 493 issues in 25 OSS projects

Most security scanning hunts for bugs. Anthropic's latest research went after something less glamorous: code that is reachable but shouldn't be. In an attack surface reduction study published October 20, Claude analyzed 25 open-source projects and more than five million lines of code, and flagged 493 issues. In OpenSSH, the fixes cut roughly a fifth of the code an unauthenticated client could reach in a default build. You don't need a frontier model to apply the lesson to your own servers.

What actually happened

Per Anthropic's write-up from its Frontier Red Team, Claude mapped reachable code across projects written in C, C++, Java, PHP, and Python, then looked for exposure no deployment needed. It used three methods:

• Default configuration: code reachable out of the box that no default feature uses.
• Configuration interaction: option pairs that are fine alone but conflict together.
• Early reachability: complex code that runs before authentication or before a protocol state change.

The numbers: more than 17,000 configuration options inventoried, 3,842 of them security-relevant. Across the gated projects, Claude found 93 unnecessary pre-authentication surfaces in default configs and 61 option combinations that silently weakened a pre-auth control. Separately, 162 security-sensitive combinations were accepted without a warning or documentation, and 94 of those silently disabled a security control.

For OpenSSH, findings went to the maintainers with patches. Release 10.5 moved a check ahead of parsing, and 10.6 closed the remaining gaps. Anthropic packaged the method as a reusable Claude skill and says it is working to integrate it into OSS-Scanner. Changes that would affect non-default configurations are left for a human to decide.

Why attack surface reduction matters for your business

Every default you don't need is exposure you don't use. The small-business server is a pile of defaults: an SSH daemon with every auth method on, a database listening on all interfaces, a CMS with XML-RPC still enabled. None of it is a bug. All of it is reachable.

Config combinations are where nobody looks. Ninety-four combinations silently turned off a security control. That is the setting someone flipped two years ago to fix a deploy, plus the one a plugin changed. Neither change looks dangerous alone.

Update OpenSSH when 10.6 reaches your distro. It's a concrete, cheap win: less code facing the internet before anyone logs in.

Use AI for the inventory, a human for the cut. Anthropic's own process leaves non-default changes to people. Let a model list what's reachable and why; have someone who knows the business decide what gets turned off.

Sources: Anthropic Frontier Red Team.