Anthropic researchers have demonstrated that frontier AI models can do more than just find vulnerabilities; they can proactively shrink the code paths attackers can reach. In a new study published on October 21, 2026, the team applied Claude to 25 large open-source projects, resulting in the discovery of 493 specific issues related to unnecessary code reachability. The core thesis is that reducing attack surface is a more scalable security strategy than hunting for individual bugs, as it mitigates the severity of both known and unknown vulnerabilities simultaneously.

The Methodology: Mapping Reachability

The research team focused on three specific methods to identify excessive exposure: default configuration analysis, configuration interaction checks, and early reachability assessment. By analyzing over five million lines of code across languages like C, C++, Java, PHP, and Python, Claude inventoried more than 17,000 configuration options. The model was tasked with identifying code that was reachable in shipped defaults but not intended for use, or code that ran before essential security gates like authentication. This approach shifts the focus from patching specific flaws to restricting access to code that simply shouldn't be exposed to untrusted inputs.

Case Study: OpenSSH and Post-Quantum Code

The most striking finding involved OpenSSH, a codebase widely considered one of the most audited in the world. Claude discovered that approximately 4,500 lines of experimental code for ML-DSA-44, a post-quantum signature algorithm, were reachable by unauthenticated clients in the default configuration of version 10.4. This represented roughly 16 percent of OpenSSH’s pre-authentication code. The issue persisted even when administrators explicitly excluded the algorithm because the daemon parsed the key type before checking if it was allowed. Anthropic worked with OpenSSH maintainers to move these checks before parsing, a fix shipped in version 10.6 that reduced the reachable attack surface by nearly one-fifth.

Configuration Conflicts and Silent Failures

Beyond raw code reachability, Claude uncovered subtle configuration traps that human auditors often miss. The model identified 162 security-sensitive configuration combinations that were accepted without error or warning, yet silently disabled critical controls. In 94 of these cases, specific option pairings turned off security features entirely. One notable pattern involved hostname-based access rules failing open when reverse DNS lookups were disabled, a behavior not documented in several projects. In another instance, a conflicting configuration pair had been present in a piece of software for 21 years, allowing unauthenticated clients to bypass certificate requirements due to a namespace mismatch in TLS settings.

Key Takeaways

  • Claude discovered 493 issues across 25 open-source projects, proving AI can identify structural security weaknesses beyond simple bug hunting.
  • The OpenSSH case study demonstrated that even highly audited code can contain significant unnecessary attack surface, with fixes reducing exposure by nearly one-fifth.
  • Configuration interactions are a major blind spot, with 162 silent security failures identified where valid settings inadvertently disabled controls.
  • Anthropic is integrating these techniques into a reusable Claude skill for OSS-Scanner, aiming to automate attack surface reduction for developers.

The Bottom Line

Finding bugs is reactive; reducing attack surface is defensive. If Claude can find 4,500 lines of unnecessary code in OpenSSH, imagine what it can do for your legacy monolith.