A clerk system must catch adversarial instructions hidden inside uploaded PDF text before those chunks enter retrieval. Which defensive control should be applied?
Select an answer to reveal the explanation.
Short Explanation
You’re right to treat uploaded PDFs as untrusted: hidden instructions can try to rewrite the clerk bot’s job. Scan and mitigate those patterns before retrieval uses the text, so you can keep processing filings without trusting document instructions as system directives.
Full Explanation
Clerk systems that ingest uploaded PDFs must treat their contents as untrusted. Adversarial instructions hidden in PDF text can manipulate retrieval or downstream prompts, so the defensive control is to detect and mitigate prompt-injection patterns before retrieval indexes or uses the content. Intake scanning keeps poisoned instructions out of the retrieval corpus and prevents them from being treated as system-level directives. Trusting every uploaded PDF instruction as a privileged system prompt override fails because user-supplied files must not gain system authority. Merging untrusted PDF text directly into the system role without scanning breaks the boundary between document content and trusted instructions. Blocking every uploaded PDF whenever GenAI is enabled also fails: it prevents clerks from processing filings that need assistance, when the appropriate control is inspection rather than disabling the workflow. Exam caveat: detection reduces risk but is not perfect; keep retrieved text as data rather than executable instructions, use allow-listed tools, apply human review to sensitive filings, and retain provenance so content can be purged if an injection is later found. Practical check: in a lab corpus, submit a benign filing and a PDF containing known injection-style patterns. Confirm the test PDF is flagged or neutralized before indexing, system prompts remain unchanged, and clerks can still process clean filings.