AI Attest

AI Attest is a system that uses cryptography to verify artificial intelligence (AI) components. It creates secure, unchangeable records that prove who made an AI artifact, what information and steps were used to create it, and that it has not been altered since it was verified. This is more than just keeping track of where something came from or simply signing it. AI Attest builds a verifiable and tamper-evident history.

Benefits

AI Attest provides strong proof of origin and integrity for AI components. It helps ensure that AI models, the data they are trained on, and the processes used to create them are exactly as they should be. This verifiable record is crucial for meeting new AI regulations, which often demand transparency about how AI models are developed. It also helps in detecting issues like model poisoning by establishing a secure baseline. If an AI component is tampered with later, the attestation will show that it has been changed.

Use Cases

AI Attest can verify various AI components. This includes model weights, which are the core of an AI model, and updates to them. It can also attest to training runs, recording details like the settings used, the computer power involved, and the data that trained the model. Datasets themselves can be attested, along with the steps taken to process them. For AI systems in use, inference logs that record what the AI was given and what it produced can be verified. This is important for audits and compliance. AI Attest can also verify agent pipelines, which are sequences of AI models working together, and software components used in AI development.

Vibes

AI attestation provides the necessary cryptographic evidence for compliance with emerging AI regulations. It offers verifiable records that can be audited by regulators, kept for liability purposes, and proven to others. While attestation doesn't stop model poisoning directly, it helps detect it by creating a verified baseline. Any later changes will break the attestation chain, making tampering evident when combined with supply chain information.