Staff paste confidential council memos into a public LLM, risking accidental leakage. Which AI risk is illustrated?
Select an answer to reveal the explanation.
Short Explanation
Dropping confidential council memos into a public LLM is classic accidental leakage — the model vendor (and maybe others) just saw city secrets. You'll still call it leakage from your pasting — you've got no hallucination making things up causing it, no prompt injection hijacking it causing it, and no training poisoning corrupting it causing it.
Full Explanation
Accidental data leakage occurs when sensitive content is submitted to systems that retain, train on, or otherwise expose it beyond authorized municipal boundaries. Staff pasting confidential council memos into a public LLM illustrates that risk: the vendor and possibly other parties may see, log, or reuse city secrets that were never approved for external processing. The harm is not malice by staff intent—it is control failure when convenient consumer endpoints become de facto document processors for confidential civic work.
Model hallucination causing accidental data leakage of confidential memos in public LLMs mislabels the stem because hallucination is about the model fabricating incorrect content, while the stem describes confidential memo text leaving municipal control through staff pasting with no fabrication causing the disclosure. Prompt injection causing accidental data leakage of confidential memos in public LLMs mislabels the stem because that risk is about adversarial instructions hijacking the model's behavior, while the stem describes staff voluntarily exposing confidential memos with no hijacking causing the disclosure. Training data poisoning causing accidental data leakage of confidential memos in public LLMs mislabels the stem because poisoning is about corrupting a model's training data to distort its future outputs, while the stem describes disclosure from pasting a confidential memo into a public LLM with no training corruption causing the leakage.
Exam caveat: distinguish the risk illustrated in the stem (accidental leakage via public LLMs) from similarly worded options that reuse the same accidental-data-leakage language but attach the wrong cause, such as hallucination causing leakage, prompt injection causing leakage, or training data poisoning causing leakage.
Operational check: block or warn on public LLM destinations from municipal endpoints, provide a sanctioned private assistant for council work, train staff on memo-handling, and monitor DLP alerts for confidential classification leaving the tenant.
Prefer to study offline? Printable PDF study book + test engine ($7.50, unofficial). These questions stay free here.