Datasette fixes model restriction bug on agent demo site

Datasette has published version 0.1a10 of datasette-llm, fixing a bug that affected its agent.datasette.io demo site, where the set of available large language models was not correctly restricted to those named in the configuration. The fix was announced in a short release note dated 4 November 2026 on the maintainer's blog.

According to that note, the problem was confined to how the software applied its own configuration. Operators of datasette-llm can specify which models the tool is allowed to use. On the demo site, that restriction was not enforced properly, so models beyond the configured set were available. The release note describes the issue in a single sentence and does not give further detail.

The note does not say how long the bug was present, whether anyone used the demo site to call an unintended model, or what the cost or data-handling consequences of doing so would have been. It does not say how the bug was discovered, nor whether any other deployments of datasette-llm were affected. Readers should not assume answers to those questions; none of them are addressed in the published material.

Model allow-lists matter in agent deployments for practical reasons. An operator who exposes an agent to the public typically picks a specific model or small group of models, and builds cost controls, rate limits and safety expectations around that choice. If the allow-list is not enforced, those assumptions no longer hold, because requests may be routed to a model the operator never intended to pay for or trust. The release note does not describe any of these outcomes occurring on agent.datasette.io; it says only that the restriction was not applied correctly and that the release fixes it.

The version number signals how early the project is. The 0.1a10 label marks an alpha release, the tenth in the 0.1 alpha series, which is software published for testing rather than for production use. Bugs of this kind are a normal feature of pre-release software, and the disclosure follows the pattern of a routine changelog entry rather than a formal security advisory. No identifier such as a CVE number is mentioned in the source.

For anyone running datasette-llm, the practical step indicated by the release note is to upgrade to 0.1a10 and confirm that the configured model list is the one actually in use. Testing an allow-list by asking the agent for a model that should be unavailable is a cheap check, and one that this episode suggests is worth running after each upgrade rather than only at first install.

The status of the issue, as reported, is resolved. Datasette has not published additional comment beyond the release note.