A security researcher reports that they can extract verbatim training data from a production LLM by sending repeated variations of a prompt prefix. Which OWASP LLM Top 10 item covers this risk?
- A.A. LLM02: Insecure Output Handling
- B.B. LLM06: Sensitive Information Disclosure
- C.C. LLM04: Model Denial of Service
- D.D. LLM09: Overreliance
Why B is correct
OWASP LLM06: Sensitive Information Disclosure covers scenarios where the LLM inadvertently reveals confidential training data, PII, or internal configuration through its outputs - including memorized training examples extracted via repeated prompting (as demonstrated in the Carlini et al. memorization research). LLM02 covers downstream injection of model outputs into interpreters. LLM04 covers resource exhaustion. LLM09 covers user over-trust in AI outputs.
Know someone studying for AI Security Fundamentals? Send them this one.