Research

Datasette Patches Security Flaws Found by Frontier AI

Datasette has released security patches 1.0a39 and 0.65.4 after developers used frontier AI models to audit the database tool and uncover subtle vulnerabilities.

Simon Willison1 day agoResearch
Illustration generated for this story

Datasette creator Simon Willison and developer Alex Garcia have released two security patch versions of the open-source data publishing tool: Datasette 1.0a39 for the current alpha series and Datasette 0.65.4 for the stable 0.65.x family. These updates address subtle security bugs discovered during an extensive audit of the codebase. The creators strongly advise administrators running Datasette instances on the public web to apply these patches immediately, particularly if their deployments mix both public and private tables.

The security push began after researcher Sevban Dönmez reported initial issues. To investigate further, Willison and Garcia conducted a comprehensive security audit leveraging several frontier artificial intelligence models, including Claude Fable 5.1, GPT-5.6, and GPT-6 Astra. The developers spent nearly a week collaborating on and reviewing the resulting fixes. According to Willison, the AI models were highly effective at identifying subtle, hard-to-find bugs in the codebase, prompting the team to commit to integrating frontier model security audits into all future development workflows.

To manage the audit findings, the developers established a highly productive collaborative workflow in a shared private repository. They split the tasks so that one developer would write automated tests to highlight a specific vulnerability, while the other would implement the corresponding fix. This methodology ensured that at least two human developers reviewed every single issue, working in tandem with their automated coding agents running the various AI models.

For software practitioners, this release highlights a growing trend of using advanced LLMs like GPT-6 Astra and Claude Fable 5.1 as active participants in security auditing. Rather than relying solely on manual reviews or traditional static analysis tools, developers can now deploy multiple competing frontier models to find edge-case vulnerabilities. Practitioners running Datasette should update their installations to 1.0a39 or 0.65.4 to secure their public-facing databases.

This is our own summary of reporting by Simon Willison

More in Research