Third-Party AI Risk

A supplier wants to embed a model in a product you depend on. The security questionnaire comes back polished, confident and thin on proof. Somebody in your third party risk function has to decide what the evidence really supports. This starter guards your data going out and the evidence coming back.

An assessment begins with a questionnaire and ends with a rating. Between those points an analyst reads security documentation, tests the claims and writes findings. That writing is now done with model help, which is exactly where an unsupported sentence enters the file.

Keep this apart from the procurement board starter. A board weighs candidates against each other and records a choice. An assessment describes a single supplier in enough detail for a risk owner to accept or refuse the exposure.

Unsupported evidence is the failure this starter aims at. Grounding checks extract each citation in a draft finding and measure it against the documentation you supplied. In the shipped settings a weak citation is flagged and the text still returns. Ask for the blocking setting when a rating depends on that source.

Assessment prompts tend to contain your architecture, your data flows and sometimes your keys. Screening for injection and a check on caller identity both run before the model is reached. Masking removes national identifier patterns. The term list refuses a request carrying the word secret, a private key or an access token.

Output scoring holds a finding to a combined 0.85, with relevance at 0.78 and accuracy at 0.80. That gate measures the writing and not the supplier. Risk appetite, tiering and contract terms stay with you. The review step remains an empty block until you name the risk owner who signs off.