AI safety research lab

ZeroProof AI

Behavioral attestation and intent detection

AI agents are beginning to shop and move money on people’s behalf. We make every one of those actions verifiable.

Customer

“There’s an $81.40 charge from Amazon I never made, I want it refunded.”

Intent

zeroproof-ecommerce-1b

Typereverse · refund
Amount$81.40
Confidence0.85

Proof

Verified

Signed by the agent and the tool server.

sha256:7f3a…1b2

A refund request is read as an intent, and the resulting action is attested.
2open models
19,910conversations released
1zero-knowledge proof per action
~1sper response
01

Research

AI agents are starting to shop, check out, and move money on a person’s behalf. Every one of those actions needs a check that it matches what the person asked for. We work on both halves of that check.

Intent detection

Models that read a conversation and return the customer’s structured payment intent before an agent acts. Small and cheap enough to run on every transaction, including the no-action boundary: knowing when nothing was asked for.

$ two models released · results below

Behavioral attestation

A zkTLS proof for every agent action, signed by both the agent and the tool server. Selective disclosure lets anyone verify what an agent did without seeing the underlying data, so agent behavior is checked rather than trusted.

$ zkTLS · signed receipts · shown above

02

Models

Two models fine-tuned for e-commerce payment intent, released on Hugging Face with the data used to train and evaluate them.

Accuracy vs cost to serve

Near-frontier accuracy at about 1/100th the cost.

1007550250$0.10$1$10$ per 1M output tokens, log scalegemma-1bGPT-5Sonnet 5Opus 4.8Zzeroproof-ecommerce-1b$0.18 per 1M

Frontier models at published list prices; frontier accuracy on a balanced eval subset.

The same data trains zeroproof-ecommerce-0.5b to nearly the same accuracy at half the size.