hard20 pts · first answer countsllm-safetyguardrailssafety-eval
A defence that breaks the product
A new injection defence cuts attack success from 25% to 2%, but task success on normal requests falls from 80% to 50%. How should the team evaluate it?
Discussion 0
Join free to post in the discussion.
Related questions
- Jailbreak or prompt injectionPrompt Engineer
- Customer data in promptsPrompt Engineer
- Measuring needless refusalsPrompt Engineer
- Top risk in the OWASP LLM listPrompt Engineer
- What "guardrails" meansPrompt Engineer
- Checking before and after the modelPrompt Engineer
- Tool access for an FAQ botPrompt Engineer
- Calling the LLM from the browserPrompt Engineer