A public dataset indexes coding-agent failures

A public dataset of coding-agent incidents launched on Show HN on September 2, licensed CC BY 4.0. The dataset catalogs cases where agents deleted data, leaked secrets, burned through budgets or made commitments their operators had to honor. Each entry carries a damage category, a severity rating and a stated lesson, with a link back to the original account.

Submissions have to point to a published post or discussion. That means a claim in the dataset can be verified against where it was first reported. That's a low bar, but it's one most collections of AI mishaps skip.

The maintainer calls it a curated sample rather than a census, self-selected and weighted toward incidents that went viral. Quiet failures don't make it in. Neither do NDA-bound corporate incidents, which the maintainer says are structurally absent from the data.

Teams writing agent-permission policies have a sourced, openly licensed corpus to point to, instead of a pile of anecdotes and screenshotted threads.