The reported issues involve LLM's extra-openai-models.yaml mechanism, which supports connections to OpenAI-compatible API endpoints from other providers. Simon Willison's Weblog says a configured OpenAI API key could be sent to another provider when the key identified by apikeyname had never been stored.
The same report describes a second scenario: a configured OpenAI API key could be sent to another provider when a model marked responses: true was run with the -o chatcompletions 1 option. Daniel de Araujo reported both security issues, according to the report.
The configuration mechanism is managed through the llm openai endpoint add command, which Simon Willison's Weblog describes as having been introduced in version 0.37. The report identifies the affected configuration as the route used for working with OpenAI-compatible endpoints from other providers.