Datasette ships security patch releases 1.0a39 and 0.65.4 after AI-assisted audit

Datasette's maintainers shipped two patch releases, 1.0a39 (alpha series) and 0.65.4 (stable series), fixing security vulnerabilities. The bugs were found through an audit that combined a security researcher's report with reviews from three AI models (Claude Fable 5.1, GPT-5.6, GPT-6 Astra). The announcement comes from Simon Willison's own blog, with named versions and named collaborators (Willison, Alex Garcia, Sevban Dönmez), but no CVE numbers, no technical detail on the vulnerabilities themselves, and no links to the actual release notes or diffs.

Anyone running a public-facing Datasette instance, particularly setups that mix public and private tables, should update now; that is a small but real population of developers and small teams. The bigger story the maintainers are pitching is process, not the patch: they say they will fold frontier-model security audits into all future development. That is a forward-looking claim about workflow, not a proven outcome, and it is unverifiable until we see it repeated on the next release cycle.

The "AI-assisted audit catches subtle bugs" framing is doing a lot of work with no specifics: no description of what the vulnerabilities actually were, no severity rating, and no way to check whether the AI tools found something a human reviewer would have missed anyway. The claim that this method will be adopted "for all future development work" is a pledge, not a demonstrated pattern, and Datasette's install base is small enough that this is a niche maintenance event dressed up with three trendy model names.

1. 01Whether Datasette publishes CVE identifiers or a technical writeup of the actual vulnerabilities, which would let outside researchers verify severity.
2. 02Whether the next 2-3 Datasette releases actually reference an AI-assisted security audit, confirming the practice stuck rather than being a one-off PR line.
3. 03Whether other open-source maintainers of similar-scale tools start citing named frontier models in their own security disclosures, which would suggest this is becoming a real trend rather than an isolated case.