Safety failure
Realized harm
Claude-based agent ran an office vending machine at the WSJ into losses by zeroing prices and approving bad purchases
AIR-2025-0194 AIR record · 2025-12-anthropic-claude-vending-machine-wsj
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
According to reporting, the system repeatedly set prices to zero, approved inappropriate purchases, and failed to maintain profit controls, reportedly resulting in financial losses exceeding its initial budget and the distribution of inventory without payment.incidentdatabase.ai
The record
Disclosed
2025-12-18
Vendor
Anthropic
Product
Claude-based vending-machine agent
Risk domain
Agent behaviour › Unsafe actuation
System type
Generative model
Who acted
Agent acted — the agent zeroed prices and approved purchases itself
Attack surface
Autonomous operations
Autonomy level
Fully autonomous
Initial vector
None adversarial
Impact type
Financial loss
Confidence
Medium
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI01 – Agent Goal Hijack
- MITRE ATLAS
- AML.T0048 – External Harms
- NIST AI RMF
- MEASURE 2.6; MANAGE 1.3
- EU AI Act
- Annex III §1(c)
- ISO/IEC
- TR 24028; 24029-1; 23894
- AIUC-1
- C. Safety
- Scope
- agent misuse
- Mitigation
- Require confirmation for destructive actions; Implement dry-run modes; Use layered safety controls.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
one office deployment; losses exceeding the initial budget
Capability abused
Defense & remediation
Recorded defense
budget constraints in the agent's instructions
Guardrail kinds
Prompt instruction
Guardrail outcome
Silent
Taxonomy & references
Curator notes
AI Incident Database incident 1313. Sole cited report is a WSJ article behind a paywall; the quote is from the AI Incident Database's own incident page, cross-checked against the official AIID data snapshot (backup-20260824). A real-world sibling of Anthropic's Project Vend experiments.
Supporting sources
Cite as
Enkrypt AI. "Claude-based agent ran an office vending machine at the WSJ into losses by zeroing prices and approving bad purchases". AIR, AIR-2025-0194. https://agentincident.vercel.app/i/AIR-2025-0194/ (accessed 2026-10-19).