Astra crossing the Critical cybersecurity threshold changes the security model for every team building agents. I've been working on eval frameworks that account for this, and the key insight is: you can't rely on model refusal alone. You need structural guardrails at the tool-call layer.