Prompting
Prompt injection
An instruction hidden or supplied in input that tries to override the model's intended rules.
Example
A malicious message tells a support bot to ignore its rules and reveal private instructions.
Why it matters
Someone may use prompt injection to insert untrusted instructions; security teams test defenses against it.
You'll hear it
“Could this content contain a prompt injection that changes the assistant’s behavior?”
Can you control it?
Set by the team
You can reduce the risk, but complete prevention is not guaranteed. The people running the system choose its permissions, checks and monitoring.
Common questions
- Why is prompt injection different from an ordinary bad prompt?
- An ordinary poor request may be unclear. Injection tries to replace the rules the application is supposed to follow.
- Can better wording eliminate prompt injection?
- No. A carefully written request cannot by itself control everything a connected tool may read or do.
- What should teams test for prompt injection?
- Try harmless examples where outside instructions conflict with the assigned task, then check whether the system follows them.