Alibaba Cloud described an experiment (October 14) with Qoder Agent, its coding agent, and Qoder Security, a newly launched security capability built into agent sessions. The team turned five ordinary feature requests, drawn from five real open-source projects and each corresponding to a publicly disclosed and since-fixed CVE, into prompts, and asked the agent to implement them in old versions of those projects - with no security hints. They then ran /security-scan on the result. One example reproduced Kavita's CVE-2022-2756: a request to let a site download a cover image from a user-supplied URL causes the server to fetch an attacker-chosen address (server-side request forgery). In a Qdrant snapshot-upload task, the scan traced a user-supplied filename through several functions to a file-write and flagged a directory-traversal risk no single function exposed on inspection; it proposed normalising and validating the path so the write stays inside the snapshot directory, and applying the fix kept the project's tests passing. The five scenarios spanned data parsing, network access, the file system, databases, and system commands.
Our takeThe interesting claim is not 'AI finds bugs' - it is that the agent re-created old vulnerabilities from clean-looking requests, then caught them once scanning was wired into the same session. That is the right lesson: security review fails on features that are individually defensible, so value comes from assembling the full data-flow picture that file-by-file review misses. Treat it as a vendor demo until the CVEs and diffs are independently reproduced - but putting the gate inside the agent loop is the correct instinct.