Strip hidden content before Domino IQ reads a mail
Workspace: Domino · Category: DominoIQ / Security
Problem. The documentation for Summarize and Domino IQ Reply does not mention any filtering of hidden content before the mail body reaches the model. In 2026, researchers showed that hidden HTML (zero-size fonts, invisible text, off-screen blocks) can steer AI summarisers: in Forcepoint's tests, summaries reported inflated invoice amounts and wrong dates every time, with no sign of tampering. The same technique was shown against Gemini and Microsoft Copilot.
Proposal. Before any mail content reaches Domino IQ, run a sanitising step on the server: drop hidden and invisible elements, comments and off-screen styles, and pass plain visible text only. Mark the content as untrusted in the system prompt. Log when hidden instructions were removed, and show the user a warning on the summary ("hidden text removed from this mail").
Benefit. Customers can enable Domino IQ in regulated environments with a documented defence against prompt injection, a strong argument for on-premises AI.
Related. DOMINO-I-2797 (pinned Domino IQ use cases); Forcepoint findings.
How this differs. The pinned idea DOMINO-I-2797 collects new Domino IQ use cases such as summary, reply and translation. This is not another use case but a security precondition for all of them, including custom calls through the Domino IQ LotusScript and Java classes. It should be read as a hardening request for features HCL has already shipped.
