AI engineering glossary · Evals guardrails AI security
What is improper output handling?
Improper output handling (also called insecure output handling): An OWASP-listed risk where an application trusts and uses model output without checking it, so the output can trigger issues such as script injection, unsafe commands or malformed data in downstream systems. Treat model output like untrusted user input.
Improper output handling: an example
A chat app inserts model text directly into a web page and an injected script tag in the reply runs in the visitor's browser.
Related terms
More Evals guardrails AI security terms
- Input guardrail
- Intrinsic hallucination
- ISO/IEC 42001
- Jailbreak
- Judge calibration
- Least privilege
- Llama Guard
- LLM-as-a-judge
Where you learn it
Improper output handling is part of the AI Engineering Program, which teaches it with notes, a hands-on lab and an exam. The first lesson is free.