Check the evidence before you send the prompt.
A hosted AI model reads your prompt in the clear. A provider that says it runs the model in a hardware trusted execution environment (TEE) makes a claim, and a claim is not evidence.
TEE Proof checks the provider’s evidence for that claim before your data goes to the provider, and stops the request when a check that is set to block fails. It is free, open source, and runs on your own machine.
The risk to your data
When your staff send documents, code or customer data to a hosted model, the provider’s machines read that data. A contract and an audit report say what the provider promises. They do not show what the provider’s machines do with your data today.
What TEE Proof does
Before it sends requests to a provider, TEE Proof gets evidence from the provider’s hardware and checks it. When a check that is set to block fails, TEE Proof does not send the request. Your data does not pass through Hardened Minds.
What you can give your counsel and your security team
- A report for each provider and model: each check, and whether it passed, failed or was skipped.
- The saved evidence. Your security team can run the checks again from it later, all but the live encryption test.
- The source code of each check, open under the AGPL-3.0, so that nobody has to take our word for it.
What it does not cover
TEE Proof does not protect against a malicious model or a backdoor in the hardware. It can check only the evidence that a provider gives. For some providers that evidence covers a gateway and not the machine that runs the model. Which checks block a request is set for each provider, and your security team can change it.
The limits, in words for a vendor-risk document
What it costs to try
TEE Proof is free. Your team installs it on one machine and points an OpenAI-compatible tool at it. Send How the checks work to your security team.