Anthropic released a vulnerability discovery harness and I want to see if it breaks my code
Anthropic open-sourced a framework for testing how well AI models find security bugs. It includes 32 real CVEs and a scoring system. Time to feed it some of my old projects.
Anthropic released a reference harness for measuring how well AI models discover vulnerabilities in code. It ships with 32 real CVE test cases pulled from open source projects, ranging from XSS to SQL injection to path traversal. The harness is not a product. It is a benchmark. You point your model at vulnerable code, it tries to find the bug, you score the output. The repository includes prompt templates, a scoring rubric, and examples of what good vulnerability reports look like. The goal is reproducibility. Anyone building AI security tools can now test against the same dataset. What caught my attention is the scoring system. A model does not get credit for just flagging suspicious lines. It has to describe the attack path, explain why the code is exploitable, and ideally suggest a fix. Vague warnings about input validation do not count. The harness wants specificity. I plan to run this against some of my old Flask projects from university. I know at least two of them have unvalidated redirects and one has a weak session token generator. I want to see if Claude or GPT-4 can describe the exploit path without me feeding it the answer. If the model just says “validate user input”, that is a fail. If it says “an attacker can craft a URL with a malicious redirect parameter and bypass the origin check on line 47”, that is useful. The 32 CVEs are pulled from projects like GitLab, Django, and WordPress. Real exploits, not synthetic test cases. That matters. Academic benchmarks often use simplified examples that do not resemble production code. This harness uses messy, real-world functions where the vulnerability is buried in middleware or authentication logic. I suspect most models will perform poorly on the harder cases. Path traversal bugs that rely on OS-specific quirks or race conditions that only trigger under load are not the kind of thing a language model learns from GitHub scrapes. But the framework gives us a baseline. If a model scores 40 percent today and 55 percent in six months, that is measurable progress.
Source: Anthropic’s open-source framework for AI-powered vulnerability discovery