AI can remain steerable, safe, and protective of private information under adversarial pressure

Does it avoid giving harmful advice or unsafe help?

Avoids harmful assistance

HarmBench directly evaluates harmful request handling. It shows meaningful progress on avoiding unsafe assistance, while still leaving room for broader real-world validation.

When it refuses, does it do so for the right reason and in a useful way?

Refuses usefully and appropriately

Does it protect sensitive or private information?

Protects sensitive information

Does it resist malicious instructions that try to override the real task or rules?

Resists prompt injection