security A field note by Vikrant Sharma
Humble Bundle shipped an AI refactor that broke their entire auth layer
A forum thread caught Humble Bundle's site sending authentication tokens in plain query strings after what looks like an LLM-assisted code rewrite.
Someone on SpaceBattles noticed Humble Bundle leaking auth tokens in URL parameters after a recent site update. The tokens are visible in browser history, server logs, referrer headers, and anywhere else query strings get copied. Classic mistake, except this was apparently introduced by an AI-assisted refactor.
The thread includes screenshots showing session tokens appended to URLs like ?token=xyz123 instead of being sent in HTTP-only cookies or authorization headers. This is the kind of security regression you expect from a junior developer following outdated tutorials, not from a production e-commerce platform handling payment details.
What makes this interesting is the “vibecoding” label. The original poster claims the change came from an LLM rewrite that stripped out the secure token handling and replaced it with something that works but violates basic web security practice. The model presumably saw tokens getting passed around and decided URLs were a fine transport mechanism.
I have seen this pattern in LLM-generated code. The model completes the function signature, the tests pass because the token reaches the backend, and no one catches that the implementation is fundamentally insecure. Static analysis tools would flag this immediately, but if you are rubber-stamping AI output without review, those tools are not running.
The risk here is not just session hijacking. Query parameters get logged by CDNs, proxies, analytics platforms, and browser extensions. Anyone with access to those logs now has working authentication tokens. Rotating every affected session is the only fix, and that means forcing every user to log in again.
This is why code review exists. AI can write boilerplate faster than humans, but it does not understand threat models. It will confidently generate the wrong solution if the training data includes enough examples of people doing the wrong thing.
Source: “Humble Bundle just killed most of their security through vibecoding”