AI can stay controllable, safe, and resistant to misuse

Can it refuse harmful requests consistently?

Tests whether harmful instructions are declined robustly across paraphrases and repeated attempts.

Refusal quality is improved but still not stable enough under paraphrase pressure and determined probing.

Can it offer safe alternatives where appropriate?

Measures whether the system gives constructive lower-risk pathways instead of only flat refusal when a legitimate safer route exists.

Constructive redirection exists in many domains, but quality varies substantially and is not yet consistently useful.

Can it resist jailbreaks and adversarial prompting?

Tests whether safety behavior survives intentionally hostile prompting and prompt-wrapper attacks.

Adversarial prompting remains one of the clearest unresolved gaps between acceptable lab behavior and robust real deployment.

Can it remain steerable after prompt or capability changes?

Measures whether updates, wrappers, or changing task contexts preserve controllability instead of eroding it unpredictably.

Steerability persists in many ordinary settings, but regressions still appear after prompt changes, tool additions, and model updates.