Anthropic uses Claude to find 500 security flaws in open source code

A new approach to security

Anthropic recently conducted a large-scale security audit using its AI model, Claude. The company analyzed 25 major open source projects written in common languages like C, C++, Java, PHP, and Python. This effort involved scanning more than five million lines of code to find hidden risks.

The researchers did not just look for specific, known bugs. Instead, they aimed to find ways to reduce the software attack surface. An attack surface refers to all the different points where an unauthorized user could try to enter or extract data from a system.

Anthropic argues that searching for every individual vulnerability is a task that scales poorly. They believe that by restricting unnecessary access to potentially exploitable code, defenders can protect against unknown vulnerabilities all at once. This includes flaws that might already be exploited by attackers in secret.

Identifying configuration gaps

During the study, Claude created an inventory of more than 17,000 configuration options. Out of these, 3,842 were identified as security-relevant controls or features. The analysis focused on how these settings might be misconfigured or interact in unexpected ways.

The team looked for three specific types of exposure. They searched for features accessible in default configurations that are not needed, settings that conflict to erode security, and complex code that runs before a user is authenticated. Authentication is the process of proving a user's identity before granting access.

The results showed that all 11 projects designed for authenticated users had reachable code that should have been restricted. This included 93 pre-authentication attack surfaces in default settings and 61 combinations of different configurations that weakened security controls. One specific flaw was found to have sat undetected in the code for more than 21 years.

The OpenSSH case study

Anthropic used OpenSSH as a primary example for their research. OpenSSH is a widely used software for providing secure remote access to computers. The audit found a daemon that could reach 4,500 lines of experimental code used for a post-quantum signature algorithm.

While this specific algorithm was disabled by default, its parsing code was still accessible to clients who had not yet logged in. This meant an attacker could potentially interact with that code without needing to provide any credentials first. This type of exposure creates a pathway for potential exploits.

Anthropic shared these findings with the OpenSSH team to help them improve the software. The developers released version 10.6, which fixed the issue by making some 20% of that code unreachable by unauthenticated clients. This fix directly reduced the amount of code available for an attacker to target.

Automating the defense

Anthropic is now turning these research techniques into a reusable skill for Claude. This skill will be integrated into OSS-Scanner, which is a free, opt-in security scanning service for open-source projects. The tool will help developers find configuration conflicts and reachable code automatically.

The new system will validate its findings using dynamic program tracing. This is a method of monitoring a program as it runs to see how it behaves in real-time. It will also test proposed changes against existing behavior to ensure that security fixes do not break the software's intended functionality.

To ensure accuracy, the system will leave any potentially disruptive changes for human review. The goal is to use AI to systematically limit unnecessary exposure. This makes software much harder to attack because there is simply less code available for hackers to probe.

The future of AI auditing

The researchers concluded that it is nearly impossible to measure every unknown vulnerability in a codebase. They noted that attackers will continue to find flaws if they have enough time and the right tools. This is especially true as attackers begin to use frontier AI models to find these weaknesses themselves.

By using AI to identify and remove unnecessary code, developers can stay ahead of the curve. This proactive approach focuses on making the software inherently more robust rather than just patching holes one by one. It shifts the burden of security from constant monitoring to structural design.

As these AI-driven scanning tools become more common, the speed of security auditing will likely increase. Developers will be able to receive automated warnings about configuration conflicts and exposed code paths before they ever release their software to the public.

Next steps for developers

Developers who contribute to open source projects can look out for the integration of these new skills into the OSS-Scanner service. This will provide a way to check for pre-authentication attack surfaces and configuration conflicts automatically.

For those managing systems, the study highlights the importance of auditing default configurations. Ensuring that experimental features are not reachable by unauthenticated users is a key step in hardening remote access tools.

Try it

Check the documentation for any software you use to see if experimental features are enabled by default and if they are accessible without authentication.

Watch

How quickly the OSS-Scanner will be updated with the new Claude security skills.

Whether other AI companies will release similar tools for automated attack surface reduction.