Summary

  • Hamachi.ai's September 10 announcement concerns a patent that its own records date to May 26, not a new September grant or independently tested privacy result.
  • Protecting a model's input does not by itself settle how the service retains context, shares data or uses de-identified material for improvement.

An advisor may want an AI assistant to forget a client's name before asking a model for help—and remember the household's circumstances next week. Those are different requirements. Hamachi.ai's privacy proposition needs to be understood across both, rather than reduced to what passes into one prompt.

The September 10 announcement promotes U.S. Patent 12,641,064 and an architecture that protects personally identifiable information before downstream LLM processing. It also describes context-driven agents, draft checking and audit logs. The timing matters: Hamachi's patent page says the patent was issued on May 26, 2026. The corresponding USPTO Gazette record identifies the patent and assignee. September is the announcement being reported, not evidence of a fresh grant or newly deployed capability.

The product story goes further than a patent number. Hamachi's homepage says sensitive data is redacted and tagged before reaching any AI model. It also markets persistent household context, advisor notes and prior decisions, and says generated communications are archived for audit. The patent page describes protecting and restoring sensitive information. These are supplier descriptions of intended operating capabilities, not an independent test of what every deployment does.

The privacy policy, effective February 17, gives the proposition a wider perimeter. It describes processing uploaded messages and integrated email content for requested features and granted permissions. For sharing with third parties, it qualifies redaction by feasibility and identifies cloud, analytics and AI model subprocessors. That broader clause and the specific pre-model claim address overlapping but different scopes. Their wording alone does not establish a leak or failed control; it does mean a buyer should not turn one input-protection claim into a promise that personal data never moves elsewhere.

Model improvement is another distinct use. The policy describes aggregated, anonymized and de-identified content being used to improve models and services. Using identifiable customer content to train a public or third-party model requires explicit authorization under the policy. This is neither a blanket no-training promise nor evidence that identifiable content is freely reused. Training must also be distinguished from processing a request.

Retention remains relevant even when a prompt is protected. The policy describes keeping data for service and other stated purposes, with some account data potentially retained after termination. Its qualified statement about generally storing summarized or derived content is not an immediate-deletion guarantee or a single retention clock for every data type.

Availability should be read just as carefully. The announcement invites private-beta access; the current site offers a trial with early-access pricing language, while the policy calls the service a commercial production platform. Those descriptions do not establish which features a particular account receives. No independent redaction benchmark, customer data-flow audit or regulatory certification is established by the cited materials.