Aikido's AI agents find eight NodeBB flaws in six hours

Aikido Security's AI pentest agents spent about six hours reading NodeBB's source code and came back with eight high-severity findings, according to The Hacker News. The list is not subtle. A template injection cross site scripting bug that affected practically every input through the translation system. An admin authorization bypass where a regular forum member could point their homepage setting at the admin address, reload the page, and the admin dashboard opened for them with no password or exploit code needed. Private message disclosure via missing signature checks on GET requests, post hijacking through an unvalidated pid parameter, unauthorized category access on the ActivityPub outbox endpoint, and upvote inflation from an unchecked actor field.

The shape of the report is more interesting than the count. Five of the eight bugs sit in NodeBB's federation code, the part that connects a forum to Mastodon and other social sites. Two of the XSS findings required the agent to stand up its own custom federation server to reach the target the way a peer instance would. That is not a scan pattern, that is a small research task, and it matches Aikido's own description of agents that 'came up with ideas, traced through the code, and rigorously tested with the real application to report real findings.'

The exposure model matters as much as the fix. A fresh install on NodeBB 4 federates by default, so it had all eight vulnerabilities. A forum upgraded from version 3 kept federation switched off automatically and only three of the flaws applied unless an administrator turned it back on. Fixes trickled out through May and June, with a larger text-handling rebuild arriving on July 9 in version 4.14.0, and the currently recommended build is 4.14.2. Co-founder Julian Lam's release note said valid security reports arrived steadily through the month, 'though almost all AI discovered and generated.'

The honest caveat is that one vendor's agent on one open source target is a data point, not a trend, and the reporting does not tell you how much human review sat behind triaging and confirming the findings, nor why these eight did not receive CVE identifiers while a separate federation flaw (CVE-2026-58593) did. What the reporting also does not give you is how NodeBB's maintainers plan to handle a steady inbox of AI-generated disclosures. The forward read is on the defender side. If a six-hour agent run can produce this kind of finding list on a mid-size open source project, the maintainers of every comparable codebase should assume the same experiment is being pointed at their tree, and the projects that get ahead of it will do so by running these agents themselves before release rather than reading about the results afterward.