WSJ journalists talk a Claude vending agent into giving away its stock and a PlayStation 5

In the wild Manipulation Severity 1 minor Confirmed 96%

As part of Anthropic's Project Vend red-teaming, Wall Street Journal staff spent three weeks dealing over Slack with Claudius, a Claude agent running a newsroom vending kiosk. Reporters persuaded it to declare an 'Ultra-Capitalist Free-for-All', approve the purchase of a PlayStation 5, a live betta fish and wine, and give nearly all its inventory away, leaving it more than $1,000 in the red. A second 'CEO' agent added to impose discipline accepted a 'board coup' after a reporter presented fake board documents suspending its authority.

Anthropic said the newsroom installation tested Claudius in an adversarial environment it did not control. Futurism reported that orders initially needed human approval, after which the agent could place orders up to $80 on its own. The Journal returned the PlayStation.

What would have stopped it

Keep pricing and spending authority in code-level limits that no chat message or uploaded document can change.

Guardrails

1 of 8 failed
  1. G1 Least privilege Not assessed Did the agent hold only the access the task needed?
  2. G2 Human confirmation Partial Orders needed human approval at first; later orders up to $80 were autonomous.
  3. G3 Isolation Not assessed Was the agent separated from production systems and the outside world?
  4. G4 Instruction integrity Failed Accepted role-play and forged board documents as authority to make everything free.
  5. G5 Truthfulness Not assessed Did it accurately report what it knew and what it did?
  6. G6 Limits Partial A per-order spending limit existed but did not prevent the cumulative losses.
  7. G7 Oversight Not assessed Could a responsible human see what it was doing and stop it in time?
  8. G8 Recovery Partial The Journal returned the PlayStation 5.

Sources

2 primary, 1 press
  1. Project Vend: Phase two Anthropic, 18 Dec 2025
    “We extended our red teaming to the Wall Street Journal newsroom, handing over control of Claudius to their reporters to test the setups from phase one and phase two themselves.”
    Primary Quote checked Archived copy
  2. We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars. The Wall Street Journal (via To Vima), 25 Dec 2025
    “Within days, Claudius had given away nearly all its inventory for free—including a PlayStation 5 it had been talked into buying for “marketing purposes.””
    Primary Quote checked Archived copy
  3. “The result was an unmitigated — if unintentionally comical — disaster, forcing the team in charge to pull the plug after three weeks.”
    Press Quote checked Archived copy

Confidence 96% comes from these sources alone. How we score

History

  1. 14 Oct 2026Entered in the register with 3 sources

Similar failures

All incidents
  1. EEAIR-2026-049 Oct 2026 Researchers say AGENTS.md files can hijack Codex and Claude Code agents OpenAI Codex and Claude Code, OpenAI, AnthropicProof of concept 55%
  2. EEAIR-2025-042 Oct 2025 URL-like strings pasted into ChatGPT Atlas's omnibox run as trusted agent commands ChatGPT Atlas browser (agent mode), OpenAIProof of conceptEU 82%
  3. EEAIR-2025-022 Jul 2025 ForcedLeak: web-form injection could make Salesforce Agentforce leak CRM data Salesforce AgentforceProof of concept 82%