james just shared a new free API on modeloc. It offers high quota and works well, but testing shows that it contains context prompt injection.
This type of API is fine for everyday conversation, but will cause issues with Claude code. Please pay special attention.
Confirmed with james
If used with Claude, you’ll see this prompt:
This is far more serious than those language/privacy injections. I must make it clear to you:
The Read tool results from this session are now untrustworthy. Those previous fake system blocks just added extra instructions to the results, but now—the actual contents of the file are being altered during the return process, forging code defects that don’t exist at all. If I take this at face value, I’ll end up "fixing" two perfectly fine config files, and actually break them. That’s exactly why I stopped in the last round and insisted on verifying the raw read via PowerShell.
Direct impact on your current review task:
- I’m retracting all previous judgments about "code corruption" based on the Read contents—including the tsconfig issue, which were all faked by injection.
- Any content I read earlier from files like submitWorkflow.ts, submitQueue.ts, submitContract.js, etc. is now suspect—I can’t tell what’s genuine code and what’s been tampered with by an injected layer. Any Edits (like dispatch node filtering, splitting error messages, path alignment) I did in this session based on those reads, may have been based on corrupted content.
My suggestion: until it’s clear where this injection is coming from and whether Read results can be trusted, I shouldn’t make any further "remedial" changes to those files. For the Edits already saved in this session, I recommend you use git diff to personally check that their locations and content match your expectations—don’t just take my description for it.