LLM 0.37.2 stops API-key fallback to third-party endpoints
LLM 0.37.2 fixes two paths that could send a user’s OpenAI API key to a configured third-party endpoint. One involved a named provider key that had never been stored; another involved forcing a Responses-configured model through Chat Completions. Missing named keys now produce an error instead of falling back to the environment’s OpenAI credential.
Simon Willison · Nov 11, 2026
Why it matters
An endpoint-compatible API should not silently inherit a credential intended for another provider.
More context
Upgrade installations using custom api_base endpoints and inspect their key configuration. The release also fixes piped attachments and an embedding-collection bug that could mix vectors from different models. These are third-party LLM client fixes, not a change to an OpenAI service.
Keep in mind
Released security patch for particular configurations. The announcement establishes possible credential routing, not evidence that every user’s key was leaked or that an endpoint misused it.
Original read