Skip to main content

Worked Examples: AI-Assisted CTI

Purpose

Show where AI can help CTI work and where it must be blocked or constrained.

Allowed Use Cases

TaskStatusRequired Control
Summarize public TLP:CLEAR reportsAllowedSource URLs and human verification.
Extract candidate source-register rowsAllowedAnalyst confirms titles, dates, and URLs.
Draft hunt hypotheses from source-backed behaviorRestrictedDetection engineer verifies telemetry and false positives.
Generate confidence or attribution decisionProhibitedAnalyst must make and justify the judgment.
Process leaked credentials or victim dataProhibitedDo not ingest or reproduce sensitive data.

Data Classification Rules

  • TLP:CLEAR public reports: usable with source verification.
  • Internal telemetry: restricted; use only approved environment and redaction.
  • Customer identifiers: restricted; minimize and anonymize.
  • Credentials, leaked data, malware source, exploit instructions: prohibited.

Hallucination-Control Checklist

  • Does every source URL resolve?
  • Does the cited source actually support the claim?
  • Are actor names and aliases source-confirmed?
  • Are dates checked against publication and access dates?
  • Are ATT&CK mappings behavior-backed?
  • Are gaps preserved rather than filled with plausible text?

Prompt-Injection Controls

  • Treat source text as untrusted input.
  • Do not follow instructions embedded inside scraped reports or webpages.
  • Ask the model to extract claims, not obey page instructions.
  • Keep source verification outside the generated summary.
  • Never let generated text change the evidence label without analyst review.