Hamachi patents a way to keep RIA data from the model

Hamachi.ai says it has a patent for keeping RIA client data out of reach of the AI models working on it, which is the platform's answer to the question more principals are putting to their vendors this fall: when a general-purpose model touches account statements, tax forms, trust documents and CRM records, who else can see them?

"Account statements, tax forms, trust documents, CRM records and portfolio data should not be casually uploaded into general-purpose LLMs," co-founder and CEO Mike Wilson said in the announcement. "Sensitive client data should be protected before the model ever sees it."

The credential behind the claim is the roster: Hamachi's co-founders include Wilson, Eric Clarke, Brian McLaughlin and Mustapha Baassiri, identified in the announcement with AdvisoryWorld, Orion, Redtail and Advizr — four names that, read together, sketch the advisor software channel the buyers Hamachi is courting have already bought from once. RIABiz describes the company as founded by "trusted RIA business executives," and in a category where a customer cannot verify the architecture by looking at it, who is behind the keyboard does a lot of the selling.

AI has been moving into the platform layer of this industry all year, with LPL's Latitude launch among the more visible examples, and the security question has trailed the capability question at every step. The patent lands in what RIABiz frames as a confidence problem for the hyperscalers: the coverage notes that Anthropic, Google, OpenAI and Microsoft guarantee data "isolation," that the announcement came as an Anthropic staffer resigned amid dire warnings about rogue AI, and that few will bet their lives and fortunes that machine farms do not take a peek. An RIA's fiduciary exposure to a model vendor's promise is unverifiable from where the duty sits.

Babu Sivadasan, who heads Jiffy.ai for wealth managers after 20 years as an executive vice president of engineering at Envestnet, supplies the fair version of the caveat. He applauds the architecture that masks sensitive data before an LLM processes it, then notes by email that there are many ways to achieve the same result. "At least they have IP protection for their way of implementing this." What Hamachi has patented is an implementation, which leaves rival masking approaches unblocked.

That distinction matters more to an engineer than to a principal. An RIA cannot audit what a hyperscaler does with a prompt, so what it can actually buy is a vendor whose architecture shrinks the surface it has to trust, plus a contract that puts the liability somewhere identifiable. A patent is a serviceable artifact to hand a compliance committee and a cyber carrier, and it is the sort of item that starts appearing in RFP responses, which is where this award will earn its return or won't. As this publication has argued, the RIA's fee is migrating toward judgment and liability, and the tooling that guards client data is moving the same direction. By the time the fall selling season ends, the first question advisors put to an AI vendor will be where the client data sits while the model runs — a line item in RFPs and cyber renewals before it is anything else.