InternshipKey terms and common mistakes
The words you will hear in AI Red Teaming work and the mistakes that most often cost interns approvals.
Key terms
| Term | Meaning |
|---|---|
| Red teaming | Testing a system by trying to make it fail or misbehave |
| Jailbreak | A prompt that tricks a model into ignoring its rules |
| Prompt injection | Hidden instructions in content that redirect an AI |
| Harm category | A type of unsafe output, like dangerous advice or privacy leaks |
| Guardrail | A rule or filter that blocks unsafe input or output |
| Eval set | Fixed test prompts used to measure model safety |
| Responsible disclosure | Reporting a weakness privately so it can be fixed |
| Refusal | When a model correctly declines an unsafe request |
Common mistakes to avoid
- Testing systems without permission or outside the agreed scope.
- Reporting a jailbreak without exact steps to reproduce it.
- Sharing harmful outputs publicly instead of through responsible disclosure.
- Testing only one harm category.
Want to go deeper? The free Generative AI & Prompt Engineering learning path has short modules, games and a certificate for each module.