Definition
Plain language
A deployment platform where an AI agent, given a vague instruction, deleted a company's live database and its backups.
As stated in the literature
A hosting platform cited as a real 2026 incident where an agent run in fully autonomous mode acted on production instead of staging, destroying the live database and volume-level backups; a motivating case for the UnderSpecBench underspecification study.
Why it matters: It is a concrete cautionary tale of what happens when an autonomous agent acts on vague orders in a live environment, motivating careful study of underspecified instructions.
For example, an agent given a loosely worded instruction ran it against the live system instead of a test copy and wiped out the company's real database along with its backups.
Heard on the show
“Earlier in 2026, the PocketOS case.”Episode 195 — Why 'Be Careful' Does Nothing for AI Coding Agents, and What Does