Source: authorized space and publisher
Process: separate text, instruction, metadata
Retrieve: access, scope, version
Execute: tools never inherit document authority
Output: citation, review, audit
Documents never gain system authority
Knowledge may contain malicious instructions, stale procedures, or copied prompts.
Use it as candidate facts while identity, permissions, and tool policy remain system-controlled.
Govern publication
Separate draft, reviewed, and published knowledge with owner, project, effective date, and version.
Preserve source access in retrieval and lower trust for external or uploaded content.
Separate facts from instructions
Flag text that asks to ignore policy, run commands, or export data as untrusted.
Represent important rules in reviewed structured definitions rather than hidden prose.
Keep least privilege at execution
Every table and tool request is rechecked against user, project, and allowlist.
Show source, version, and excerpt, and require review for high-impact conclusions.
Accept with injection samples
Test hidden instructions, conflicting rules, stale versions, cross-project references, and links.
AskTable.ai supports documents and semantic configuration directions; review, isolation, detection, and citation need project verification.
Public references
Ready to help your team start?
Talk through a real scenario and see how AskTable.ai can fit your business.
Book a demo