Proof of Provenance for Web Retrieval

As web-search-augmented LLM prompting becomes more commonplace, one issue that is likely to arise is that the web context that the LLM is using has to be trusted not to be manipulated. One solution for this is to add “proof of provenance” for that data, such that the integrity of the data retrieved from a third-party website is verifiable.

As an example case, an agent which is relying upon web search for information that it will use to make decisions might want to ensure that nobody is tampering with its information sources.

The simplest, future form of this will be RFC-9421, which creates a standard for how web services can sign their responses. However, development and adoption of that standard is likely to take several years at least.

In the meantime, an existing solution is MPC-TLS, particularly TLS Notary. The tl;dr is that the web retrieval service for the LLM would execute its request over a connection constructed in collaboration with the requesting client (which acts as a notary), or with a third-party notary that the user trusts not to conspire against them (e.g. running in a TEE).

With this, the response to the client can include the proofs showing that the web content is untampered.