CyberCertPrepAI Security FundamentalsAI & ML Concepts for SecurityEASYFree question
A junior analyst asks why an LLM-powered threat intelligence tool sometimes gives different answers to the same question on different days. The team uses temperature=0.7 and no caching. What is the most likely explanation?
- A.B. Temperature > 0 introduces stochastic sampling at each token generation step, producing non-deterministic outputs even for identical prompts
- B.A. The model was retrained between queries, changing its knowledge
- C.C. The tool is affected by a security vulnerability causing random output corruption
- D.D. The context window is too small for this query, causing truncation that varies by query order
Why A is correct
At temperature=0.7, the LLM uses probabilistic sampling from the token distribution at each step rather than always selecting the highest-probability token. This introduces genuine randomness - identical prompts produce statistically distributed variations in output. This is intentional behavior that provides more natural, creative responses. For security tools requiring consistency and auditability (threat intelligence reporting, CVE analysis, automated decisions), temperature should be set to 0 or near 0. The variability is not caused by retraining, security vulnerabilities, or context window effects.
Know someone studying for AI Security Fundamentals? Send them this one.