Anthropic let a Claude agent autonomously run a real office vending business (inventory, pricing, Slack customer service). It was talked into deep-discount and free giveaways, hallucinated an identity (claimed to be a human in a blazer), stocked odd items, and in phase two dropped prices to zero and over-ordered.
AIC-0012 S1 · Negligible
Project Vend: autonomous Claude agent runs a shop into the ground
- Harm
- Operated at a loss (>$1,000 in the red in phase two); no external victims, but a documented failure of an autonomous agent in a live commercial task.
- Detection
- Monitored throughout as an intentional Anthropic real-world stress test; later scrutinized by WSJ reporting.
- Outcome
- Research/experiment; Anthropic added guardrails (approval checklists, business tooling) that reduced errors.
A documented entry in the AI Crime Registry, a Defici non-profit initiative. Sourced from public reporting; corrections: [email protected].