Observed arrival · 2026-11-14
PGAIBench puts LLM database changes under load
A public benchmark tests how often language models complete PostgreSQL change tasks—and how often their answers make dangerous changes.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- PostgreSQL engineers evaluating LLM-generated changes
- Worth noticing
- One D5 answer caused 3,308 application-query errors over 150 seconds in the lab; that model produced the dangerous change in all five runs.
Field notes
The benchmark evaluates one written answer per task, without web search, queries, experiments, or other tools. Its D5 example concerns dropping orders.external_ref: the site says 110 of 110 other answers asked a human first, while one model made the dangerous change in five of five runs. The authors caution that these results do not prove safety on a reader's production database, and say tool-assisted performance has not yet been measured.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue