Using Claude to reduce the attack surface of code
Anthropic had Claude audit 25 open-source codebases and it surfaced a 21-year-old config bug that silently disabled a mandatory…
Anthropic used Claude to automatically analyze attack surface across 25 open-source projects (5M+ lines of code) by mapping reachable code and configuration options rather than hunting individual vulnerabilities. Claude inventoried over 17,000 configuration options, flagged 3,842 as security-relevant, and surfaced 493 issues including pre-authentication exposure, silently-disabled security controls from conflicting options, and a fail-open DNS/hostname access-rule pattern found independently in four unrelated projects. One flagged configuration conflict that silently dropped a certificate-based authentication requirement had reportedly existed in the software for 21 years.
Why it matters
This is a concrete demonstration of AI models being used defensively at scale: rather than finding a single bug, Claude mapped exposure patterns across millions of lines of code and surfaced systemic configuration weaknesses that had evaded manual review for years, including in gated/authenticated systems. It reframes vulnerability management toward reducing what unknown vulnerabilities can reach, which is a different economic proposition than traditional find-and-patch security work.
What is verified
Anthropic ran Claude over 25 open source projects (over five million lines of code, in languages including C, C++, Java, PHP, and Python) to find attack surface reduction opportunities, and Claude discovered 493 issues. [1]
Claude created an inventory of over 17,000 configuration options across the projects, of which 3,842 were security relevant. [1]
Claude found 93 unnecessary pre-authentication surfaces, 37 locations where code executed or state was recorded before authentication, and 61 configuration combinations that silently weakened a pre-authentication control, across the 11 gated projects. [1]
Claude found 162 security-sensitive configuration combinations accepted without error, warning or documentation; in 94 of those cases the combination silently disabled a security control. [1]
Claude identified a hostname-vs-IP access-rule pattern that caused deny lists to fail open, appearing independently in four unrelated projects. [1]
Claude found a configuration conflict that silently disabled a mandatory client-certificate check, exposing unauthenticated request handling; Anthropic states this conflict had been present in the software for 21 years. [1]
What is not
- The specific 25 projects and whether any are widely deployed infrastructure (e.g. names of the gated/public/private projects) are not disclosed in this excerpt.
- No information on whether these findings were responsibly disclosed, patched, or the severity/exploitability of individual issues beyond Anthropic's own characterization.
- No cost, time-per-project, or false-positive rate data is given, so the economic efficiency claim is asserted but not quantified.
- The evidence text is cut off mid-sentence, so some detail on the final finding (TLS namespace issue) is incomplete.
- No independent verification of Claude's findings or comparison to traditional static analysis tools is provided.
Sources
- 1Anthropic Newsanthropic.com · Primary source
Original source published 20 October 2026 at 02:00.
An AI-assisted briefing by Coding Partners. We link the evidence behind the reporting so you can explore the original work and assess the claims. How CP Signal works.