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

The scale of the audit

Anthropic recently employed its Claude AI model to perform a massive security audit of the open source ecosystem. The company examined 25 major projects written in several common programming languages, including C, C++, Java, PHP, and Python. This analysis covered a total of more than five million lines of code.

During this process, Claude created a comprehensive inventory of over 17,000 configuration options. From this list, the AI identified 3,842 specific options that were relevant to security controls and features. The goal was to see how these settings might leave software exposed to hackers.

The audit revealed nearly 500 security issues across the projects. These included instances where security controls were turned off due to conflicting settings and areas where sensitive code was accessible before a user could even log in. As TechRadar reported, the scan even uncovered a vulnerability that had been sitting in the code for over two decades.

A shift in security strategy

Anthropic argues that the current method of hunting for individual vulnerabilities is a Sisyphean task. This means the work is repetitive and never-ending because it is impossible to guarantee that every single bug has been found. The researchers suggest that this approach scales poorly as software becomes more complex.

Instead of just looking for specific bugs, Anthropic advocates for reducing the software attack surface. This involves restricting unnecessary access to any code that could potentially be exploited. By making less code reachable by outsiders, developers can protect against all unknown vulnerabilities in that specific area simultaneously.

The company believes this strategy is more effective because it addresses the root cause of exposure. Anthropic noted that reducing the attack surface minimizes the risk from all unknown vulnerabilities in that code at once. This approach helps mitigate risks from flaws that might already be being quietly exploited by attackers.

Identifying configuration gaps

The research focused on three specific ways that code becomes unnecessarily exposed to the public. First, it looked at features that are active in default configurations even when they are not needed. Second, it examined different settings that, when combined, can accidentally erode existing security protections.

The analysis found that all 11 projects designed for authenticated users contained code reachable when it should not have been. This included 93 pre-authentication attack surfaces in default setups and 37 instances of code executing before a user provides credentials. It also found 61 combinations of settings that weakened these pre-authentication controls.

In some cases, these configuration issues were so severe that security controls were completely disabled. One specific flaw involved a setting that required certificates for identity but another setting that disabled that requirement entirely. Because the software did not warn the user about these conflicting options, the security gap remained open.

Case study: OpenSSH

Anthropic highlighted OpenSSH, a widely used tool for secure remote access, as a primary case study. The software contained a background process that allowed access to 4,500 lines of experimental code. This code implemented a post-quantum signature algorithm but was accessible to unauthenticated clients before the server could check if it was permitted.

Even though the algorithm was turned off by default, the fact that the parsing code was reachable created a risk. Anthropic disclosed these findings to the OpenSSH team, who developed a fix in version 10.6. This update ensured that approximately 20% of the code was no longer reachable by unauthenticated clients.

This success story demonstrates how AI can identify specific areas for cleanup. By removing the reachable experimental code, the developers significantly narrowed the window of opportunity for potential attackers. This type of systematic reduction is what Anthropic wants to scale across the broader open source community.

Automating the defense

Anthropic is now turning these research findings into a reusable skill for Claude. This skill will soon be integrated into OSS-Scanner, which is a free and opt-in security scanning service for open-source projects. The tool will automatically examine default configurations and identify settings that are in conflict with one another.

The system will also look for code that is reachable before security checks are completed. To ensure accuracy, the system will validate its findings using dynamic program tracing. This technique allows the AI to see how the code actually behaves when it is running.

The AI will also test proposed changes against existing behavior to ensure the software still works as intended. However, Anthropic noted that any changes that could be potentially disruptive will be left for human developers to review. The ultimate goal is to use AI to help developers systematically limit unnecessary exposure, making software much harder to attack.