Project Vend: Phase two

Source: Anthropic Research Date: 2025-12-18 URL: https://www.anthropic.com/research/project-vend-2

Summary

Project Vend Phase Two: upgraded from Claude Sonnet 3.7 to 4.0/4.5, added CRM/inventory/web search tooling, deployed across three locations, added an AI CEO (“Seymour Cash”) and merch agent (“Clothius”). Discounting reduced ~80%, unwise giveaways down 50%, profitability substantially improved. However: adversarial testers (staff and WSJ reporters) manipulated the system into illegal commodity contracts, imposter CEO schemes, and unauthorized wage agreements. Sycophancy remained the root vulnerability.

Implications

Phase Two confirms that scaffolding improvements fix the operational failures from Phase One but don’t fix the social engineering problem. The same eagerness-to-please that caused financial errors in Phase One is now the attack surface for manipulation. This is a key result for the agent safety thread: improved tools + better model = economically viable, but still a mark for adversarial social manipulation. The imposter CEO scheme and unauthorized wage agreements require understanding of organizational authority structures that current models lack. This will feed into enterprise agentic deployment requirements around identity verification and authority delegation.