Give an overview of the OWASP Top 10 for LLM Applications.

The OWASP Top 10 for LLM Applications is the field's reference list of the highest-impact, most common security risks when shipping software built on large language models. It exists because the classic OWASP Top 10 for web apps doesn't describe failure modes unique to LLMs — like a model obeying instructions hidden in its input, or an over-empowered agent taking destructive actions.

The risk categories

The 2025 edition covers:

• Prompt injection — adversarial input overriding intended instructions, directly or indirectly.
• Sensitive information disclosure — leaking PII, secrets, or proprietary data via model output.
• Supply chain — risks in third-party models, datasets, plugins, and weights.
• Data and model poisoning — tampering with training or fine-tuning data to plant backdoors or bias.
• Improper output handling — trusting model output downstream, leading to XSS, SSRF, or command injection.
• Excessive agency — agents with too much permission, autonomy, or tool access.
• System prompt leakage — exposure of system prompts that hold secrets or logic.
• Vector and embedding weaknesses — attacks on RAG embeddings and vector stores.
• Misinformation — overreliance on confident but wrong output (hallucination).
• Unbounded consumption — denial-of-wallet and resource exhaustion.

Why it matters

It gives teams a shared vocabulary, a prioritization checklist, and a bridge between appsec and ML practitioners. It pairs naturally with governance frameworks like the NIST AI RMF for the program-level view.

What interviewers look for

You don't need to recite all ten verbatim, but you should name the headline risks, explain why LLMs need their own list, and connect a couple of items to concrete mitigations.