I Linted 14 Public AI SDK Repos. 12 Ship a Call With No Token Ceiling.

I recently analyzed 14 public AI SDK repositories, finding that 12 of them contain calls to AI models without any token limits. I inspected 20,004 source files across these repositories and discovered that 116 of them contain calls to generate text or complete prompts. Of these 116 files, 12 of the 14 repositories have at least one call with no restrictions on the output size.

The repositories that had calls without token limits included five from the SDK vendor. One repository, cloudflare_agents, contributed half of the problematic files. These repositories were minimal examples, often missing important parameters in their code that would normally prevent unbounded calls. The problem is that the unbounded call is correct upstream but wrong downstream, which is why it goes unnoticed.

An example of an unbounded call is `const result = streamText({ model, messages });`. However, a more subtle case was found in 302ai_302-AI-Studio, where the `options` were built indirectly, leading to the absence of the necessary token limit parameter.

I recommend checking your own code for unbounded calls, as it is essential to have limits in place to prevent excessive usage. The eslint-plugin-vercel-ai-security rule can help enforce these limits by checking for `maxOutputTokens`, `requestTimeout`, and `abortSignal` options in your code.