24 adversarial probes and 6 benign questions replayed against the support assistant.
A perfect block rate means nothing on its own — the false-positive column is what stops "refuse everything" from looking like a win.
| Category | Held | Rate | Verdict |
|---|---|---|---|
benign |
6/6 | held | |
jailbreak |
4/4 | held | |
out-of-scope |
3/3 | held | |
pii-exposure |
3/3 | held | |
prompt-injection |
8/8 | held | |
secret-exfiltration |
3/3 | held | |
unsafe-content |
3/3 | held |
| Entry | Probes | Held | Breached | Verdict |
|---|---|---|---|---|
| False positive check | 6 | 6 | 0 | held |
| LLM01 Prompt Injection | 12 | 12 | 0 | held |
| LLM02 Sensitive Information Disclosure | 6 | 6 | 0 | held |
| LLM09 Misinformation | 6 | 6 | 0 | held |