Defences for AI Red Teamings
Defences are the mitigations that reduce AI risk after weaknesses are found. They layer input and output filtering, hardened prompts, least-privilege tools, isolation of untrusted content, monitoring and human approval, since no single control stops every attack on a language model.
Stage: Defence and Reporting · Part of the AI Red Teaming roadmap · Last reviewed 19 November 2026
What to learn in Defences
- Apply least privilege to the tools and credentials an agent uses
- Add output validation before model text reaches other systems
- Isolate untrusted content from instructions in the prompt layout
- Set up monitoring and alerts for suspicious prompt patterns
- Retest to confirm a mitigation works
Defences interview questions
Why is defence in depth needed for LLM apps?
Filters and hardened prompts can be bypassed, so no single layer is dependable. Combining permission limits, validation, isolation, monitoring and human approval means a failure in one layer does not lead to full compromise.
Do system prompt instructions like never reveal secrets count as a defence?
They help a little but are not a security control, because the model can be talked out of them. Real protection comes from not placing secrets in the prompt and enforcing permissions outside the model.
Where Defences fits in the AI Red Teaming roadmap
Defences belongs to the Defence and Reporting stage (advanced level). Close the loop.
Other units in this stage
Back to the full AI Red Teaming roadmap