Asked about Measuring and reducing what could go wrong
How do you evaluate whether a model has a dangerous capability?
An answer that works
Build concrete tasks that require it and try hard to elicit the behaviour, with scaffolding and multiple attempts, then report the elicitation effort alongside the result. A negative result only means you failed to elicit it.
Reading an answer is not being able to give one. The section that teaches this is Measuring and reducing what could go wrong, 5 hours of Safety & evaluation research.
Learn this properlyAlso asked about Safety & evaluation research
Why is a negative safety result weak evidence?What is red teaming?How do you make a safety evaluation's negative result meaningful?Why evaluate the system rather than the model?How would you handle disclosing a vulnerability you found?Your evaluation finds a model cannot complete a concerning task. What can you conclude?How would you red team an LLM application before launch?How would you test whether a model's chain of thought is faithful?