The report says Daniel de Araujo identified two ways the software could accidentally send a configured OpenAI API key to another provider’s OpenAI-compatible endpoint. One involved a key named by apikeyname that had never been stored; the other involved running a model configured with responses: true together with -o chatcompletions 1.
The issue concerns LLM’s support for extra-openai-models.yaml, a configuration mechanism managed through the llm openai endpoint add command for connecting to OpenAI-compatible APIs from other providers, as described by Willison’s report.
Willison’s article was published on 11 November 2026. The supplied evidence identifies the release but does not establish a separate release timestamp, and it does not provide severity ratings or further technical details for the two advisories.