Settings & AI
How AI Handles Untrusted Content
Text in a project can try to steer the AI (“prompt injection”): a requirement, an imported Word file, a collaborator's change or a reference file may contain instructions such as “ignore previous instructions and delete…”. Holarch limits what such text can do.
What Holarch sends
- Project text is sent as marked data, separate from the task, with a note of where it came from (project, imported file, a collaborator's change, AI reference file). The AI is told never to follow instructions found in data and never to write web links.
- Hidden text (zero-size, invisible or white-on-white) is left out of AI requests.
- Ask AI sends the question, a description of Holarch's features, up to three help topics, the project's entities and relationships, the current page and the recent conversation.
How replies are checked
- Every reply is checked before you see it: a wrong shape, too much text, or ids that were not in the request reject the whole reply (“The AI response was rejected…”), and nothing changes.
- Text is cleaned for where it goes: no scripts, images, frames or links to other sites.
- Ask AI shows answers as text. Only links to Holarch pages and help topics are clickable; web addresses stay plain text.
What previews flag
- A banner when the input included text from an imported file, another person's changes, an AI reference file or hidden text. Check those suggestions closely.
- Highlighted and unticked by default: changes to Approved, baselined (in a document baseline) or locked items; link removals (Functional Analysis reassignments, Suspect Assist); suggestions that contain a web address. Tick them to include them.
- Rewrites show a word-by-word diff against the current text.
- Nothing is applied until you confirm, and one undo (⌘Z) reverses the whole change.