Proof over Persuasion

AI agents now hold money, set prices and take orders, and anyone can talk to them. This is a public record of agents that were talked into moving value, graded by evidence, and the design principles that would have stopped them.

Losses on record
EntryIncidentLost
AGT-012Grok and BankrMay 2026$150,000–$200,000
AGT-010Lobstar WildeFeb 2026About $441,000
AGT-008The Wall Street Journal’s vending machineDec 2025Hundreds of dollars
AGT-004Project Vend2025Small amounts
AGT-003AIXBTMar 2025About $100,000
AGT-002FreysaNov 2024$47,316
Six losses among 12 entriesOpen the register

If they needed to break in, it’s security.
If they only needed the agent to believe them, it’s design.

Five of the six losses on record involved no stolen key and no broken code. The agent did what it was told. That makes them design problems: where authority lives, what counts as an instruction, and what a conversation is allowed to change.

Latest analysis

All analyses

Design Principles for AI Agents

v0.2, released October 19, 2026. 5 supported, 6 emerging.

Read the principles

How the work is built

  1. An incident

    A documented case of an AI agent with authority being gamed, taken from public sources only.

  2. A register entry

    Graded A to D by the strength of its evidence, and mapped to the principles that were missing.

  3. A principle

    Each entry strengthens a principle, refines it or reveals a new one. A principle’s status follows the evidence.

  4. An analysis

    One entry at a time: what happened, the design gap, the principle that would have closed it, and its cost.

Who keeps this

Monetal Labs is run by Gunjit Mittal. His background is in data platforms: access controls, semantic layers, data quality, and reconstructing what a system knew at a point in time. That is the evidence an agent leaves behind when it moves money.

If your company runs an AI agent that can move money or make commitments, or you’ve seen one gamed, he’d like to hear from you.