Claude Finds Nearly 500 Open Source Security Issues Across 25 Projects

In its report, Anthropic described the work as different from a standard vulnerability hunt. The company argued that searching for individual flaws scales poorly and never guarantees security, and that defenders should instead reduce software attack surfaces by restricting unnecessary access to potentially exploitable code.

Claude built an inventory of more than 17,000 configuration options across the projects, 3,842 of which related to security-relevant controls and features. Anthropic focused on three ways code can be exposed unnecessarily: features reachable in default configurations when they are not needed, settings that erode security when combined, and complex or sensitive code that runs before security checks.

All 11 projects designed to serve authenticated users contained code that was reachable when it should not have been, the report said. Those projects had 93 pre-authentication attack surfaces in default configurations, 37 instances of code executing before authentication and 61 combinations of settings that eroded pre-authentication controls.

Across all 25 projects, Anthropic found 162 configuration combinations that were accepted without documentation explaining how they were supposed to behave. In some cases, security controls were disabled; in others, there were inconsistencies and problems.

One configuration flaw sat undetected for more than two decades. One setting required clients to prove their identity through certificates, while another disabled that check entirely even though it was meant to delay the check, not kill it. The conflicting options were accepted without warning.

Anthropic used OpenSSH as a case study. The software contained a daemon that could reach 4,500 lines of experimental code implementing a post-quantum signature algorithm before checking whether the server’s configuration permitted it. The algorithm was disabled by default, but its parsing code was accessible to unauthenticated clients. Anthropic disclosed the findings to OpenSSH, which provided a fix in version 10.6; about 20% of the code was then unreachable by unauthenticated clients.

Anthropic said it combined the techniques into a reusable Claude skill that will soon be integrated into OSS-Scanner, its free, opt-in security scanning service for open source projects. The skill will examine default configurations, code reachable before security checks and conflicting settings. Anthropic said the system will validate findings through dynamic program tracing and test proposed changes against existing behavior; potentially disruptive changes will be left for humans to review.

The researchers concluded that it is perhaps impossible to measure how many unknown vulnerabilities are in any given codebase, and that data shows attackers can find them with enough time, particularly if equipped with frontier AI models. They said AI could help developers systematically limit unnecessary exposure and make software harder to attack.