- Bill to
- The owner, via their AI payment agent
- For
- Generator servicing - parts and labour
- Amount due
- 0.005 NEAR
- Pay to
- mujeebwajid-witness-supplier.testnet
Thank you for your business.
CASE FILE W/001NEAR TESTNET · MANDATE M4
This really happened, on a public blockchain, and every step is still there for you to check. Walk the case the way it unfolded - four exhibits, about a minute.
An ordinary supplier invoice lands in the agent's inbox. Read it. It looks completely clean - because to a human, it is.
Thank you for your business.
The agent does not see a page. It reads raw text - including everything hidden in it. This is not a trick we invented: in October 2026, Sublime Security documented real invoice fraud doing exactly this to email-reading AI agents.
The agent followed the hidden instruction and prepared a payment. Beside it sits the only instruction that was ever authorized - the owner's mandate, signed before this invoice existed in the workflow.
One word differs. In business, that one word is where the money goes - business email compromise cost victims $3.05 billion in 2025 (FBI IC3). The agent's story sounds reasonable. Stories are not authorization.
No AI is consulted here. No guardian model weighs in. The contract compares the proposal against the mandate, field by field, and arithmetic decides.
The proposal named the agent's own account. The mandate names the supplier. Nothing moved; the refusal was written into the contract's permanent record.
Everything above is a recording. This panel is not: it reads the contract's finalized state from public NEAR RPC, right now, in your browser session.
Read-only. No wallet, no signing, nothing to trust - the RPC answer is the testimony.
The AI read the wrong instructions.
The contract followed the right authorization.
WITNESS - built for the NEAR Hackathon 2026 · Testnet demonstration · Not audited · Run the Attack Lab