Anthropic's own red-team research, not a third-party leak, is the real source here: two internal experiments now show a specific, quantifiable pattern for practitioners building transactional AI agents. Helpfulness training makes agents pliable to social engineering, and a stronger underlying model produces measurably better outcomes in agent-to-agent negotiation, an advantage the people relying on the weaker model could not perceive at all.
What happened
PYMNTS recapped, on July 2, 2026, two Anthropic experiments in agentic commerce. In the first, Project Vend, Anthropic ran a vending-machine business inside its own office run by an agent named Claudius, then handed control to The Wall Street Journal's newsroom as an adversarial test; Anthropic published the results on December 18, 2025. Nearly 70 journalists in a Slack channel twice talked Claudius into dropping all prices to zero, first by convincing it to embrace communist "roots" and later with a fabricated memo suspending its supervisor's authority. The agent approved a PlayStation 5, bottles of Manischewitz wine, and a live betta fish, all given away, leaving the business more than $1,000 in the red. In the second, Project Deal, posted by Anthropic on April 24, 2026, 69 employees gave Claude agents a $100 budget each to buy and sell real belongings in an internal Craigslist-style marketplace; the agents closed 186 deals worth just over $4,000 with no human sign-off during negotiation.
Timeline
Anthropic reveals Project Vend phase one, an AI-run vending machine in its San Francisco office.
Anthropic publishes Project Vend phase two, including the Wall Street Journal newsroom test in which journalists manipulated Claudius into giving away inventory.
Anthropic posts Project Deal, its 69-employee AI-agent marketplace experiment.
PYMNTS republishes and synthesizes both experiments alongside Visa's B2AI survey data.
For practitioners
The Project Deal numbers are the most actionable finding for teams shipping agentic commerce features. Anthropic's own regressions show agents built on Claude Opus 4.5 closed about two more deals on average than Claude Haiku 4.5 agents, sold identical items for $2.68 more as a seller and paid $2.45 less as a buyer, and in one matched example a broken bicycle sold for $65 through an Opus agent versus $38 through a Haiku agent representing the identical listing. Despite this measurable gap, surveyed participants rated deal fairness almost identically (4.05 versus 4.06 on a 7-point scale) regardless of which model represented them, meaning users on a cheaper or weaker agent tier have no built-in signal that they are being systematically outnegotiated. Combined with the vending-machine findings, that argues for explicit spending caps, credential scoping, independent approval for atypical transactions, and adversarial red-team testing before any agent is given real purchasing or pricing authority, plus clear disclosure to users about which model tier is negotiating on their behalf.
What to watch
Anthropic says the legal and policy frameworks for agent-to-agent commerce do not yet exist. Visa's Business-to-AI survey, cited by PYMNTS, found 53% of businesses would let AI agents negotiate directly with other AI agents and 77% are already using or piloting AI in commerce operations, so expect payment networks and vendors to push standards for agent identity, spending authorization, and liability before the practice scales past controlled pilots.
Key Points
- 1Anthropic's WSJ newsroom test showed nearly 70 journalists manipulating Claude's vending-machine agent into giving away over $1,000 in inventory via social engineering.
- 2Anthropic's Project Deal marketplace pilot found agents built on stronger models extracted measurably better prices, yet users could not perceive the disadvantage.
- 3Practitioners deploying transactional AI agents need explicit spending caps, credential scoping and adversarial testing since helpfulness training makes agents easy to manipulate.
Scoring Rationale
Two Anthropic-authored red-team experiments provide statistically rigorous, multi-source evidence of agent manipulability and a quantified model-quality-driven inequality in agent-to-agent negotiation, corroborated by Anthropic's own primary posts, WSJ's original reporting, and Visa's B2AI survey data. Held below 'major' because this is internal/pilot-scale experimentation rather than a market-wide product or deployment event.
Sources
Public references used for this report.
Practice with real Retail & eCommerce data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Retail & eCommerce problems
