llm 0.37.2
Simon Willison released llm 0.37.2, a patch to his command-line tool for running prompts against large language models. It closes a security issue reported by Daniel de Araujo and tracked in two advisories, GHSA-v9h7-jj4g-f862 and GHSA-r639-rr7r-x72w.
The exposed surface is extra-openai-models.yaml, the config file that points LLM at OpenAI-compatible APIs run by other providers. That file is now managed by llm openai endpoint add, a command introduced in 0.37. The risk is credential leakage: your configured OpenAI API key could be transmitted to one of those third-party providers.
The release note describes two routes. The first is a reference to a key by api_key_name where that key had never been stored. The second is running a model marked responses: true with the option -o chat_completions 1, which forces the older chat completions path. Both caused the configured OpenAI key to be sent onward rather than withholding it.
In practice, anyone who has added third-party endpoints and either misnamed a key or overrode the responses flag on a responses-based model should upgrade and treat the OpenAI key as potentially exposed. Rotating it would be prudent.
Why it matters: developers using LLM with non-OpenAI endpoints were exposed to quiet key leakage, and a forced fallback flag is an easy mistake to make; patching and rotating credentials is the safe response.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Release: llm 0.37.2 Fixes a security issue reported by Daniel de Araujo in GHSA-v9h7-jj4g-f862 and GHSA-r639-rr7r-x72w . LLM has a mechanism called extra-openai-models.yaml - now managed by the new-in-0.37 llm openai endpoint add command - for using LLM with OpenAI-compatible API endpoints from other providers. Daniel reported two ways this could accidentally send your configured OpenAI API key to one of those providers: if the key named by api_key_name had never been stored, or if a responses: true model was run w