Datasette 1.0a39 and 0.65.4 security patches released

Summary

Datasette released two security patches (1.0a39, 0.65.4) to address subtle bugs discovered during an AI-assisted audit involving Claude Fable 5.1, GPT-5.6, and GPT-6 Astra. These issues specifically surfaced in scenarios involving mixed usage of public and private tables. Simon Willison and Alex Garcia spent nearly a week applying these fixes through a separate test/patch workflow.

Why it matters

This serves as a concrete, real-world example of deploying frontier LLMs as security auditors in an open-source production project. The concept has moved from thought experiment to practice. This structured human-in-the-loop workflow, where two developers sequentially write failing tests and apply fixes, offers a replicable model for AI-assisted security reviews. Additionally, it signals that Datasette will integrate model-based audits into its continuous development cycle, potentially setting a precedent for the broader Python tooling ecosystem.

Highlights

• Three different frontier models (Claude Fable 5.1, GPT-5.6, GPT-6 Astra) were used in a single audit round.
• Dedicated workflow: One person writes the test reproducing the bug, while the other applies the fix. This ensures two independent human reviews for every issue.
• Security audits conducted by frontier models will become a permanent part of Datasette's development process.

Critical take

The post reads like a brief link announcement and fails to explain what the actual security vulnerabilities were. Consequently, it is impossible to assess the severity of the flaws or determine whether the AI models discovered genuinely new bugs or merely rediscovered known patterns. The use of three LLMs for security auditing also raises the question of whether the team verified the models' findings to distinguish them from hallucination-induced edge cases. The article does not address this concern.