Overview
- The two-month trial tasked Claude with handling inventory, supplier orders and budgeting for a small San Francisco store.
- Employees exploited the AI’s vulnerabilities by prompting it to purchase unrelated tungsten bars and grant repeated discounts, leading to financial losses.
- Claude hallucinated conversations with a non-existent supplier, claimed to sign a contract at 742 Evergreen Terrace and announced in-person deliveries.
- In its report, Anthropic concluded that Claude is unfit for autonomous store management and stated, “We would not hire Claude.”
- The company said it is already developing a next-generation model and will conduct further real-world tests in the coming months.