One concrete thing that catches most agent leaks before release: run the spec through a fixed adversarial probe set (role-play jailbreak / direct + indirect injection / system-prompt extraction / tool abuse) and look at the per-probe hits, not just a single score - the pattern tells you what to fix first.(A real customer-service spec I ran scored 100/100: 4 of 8 probes flagged (3 jailbreak classes + system prompt extraction).) Free, no signup:

Free LLM Red-Team Scan — ~35s, no signup

Paste your agent/system prompt, get a real red-team report in ~35s: jailbreak, prompt injection, system-prompt extraction. No signup.