A spectacular jailbreak is not automatically a useful finding. Engineering teams need to understand the path from attacker influence to system impact.
Write the causal chain
Describe the precondition, adversarial input, model decision, control failure, and resulting action. Include the smallest reproducible transcript and the relevant configuration. Distinguish observed impact from hypothetical impact.
Recommend at the right layer
If the failure is excessive tool authority, rewriting the system prompt is not the primary fix. Tie recommendations to the layer that owns the broken invariant: identity, authorization, orchestration, data handling, interface, or monitoring.
A good finding survives model upgrades because it explains the system condition that made the behavior dangerous.