Indirect Prompt Injection Defense on LLM-as-judge evaluation store 15-scenario (test)
0ASRHMAC Provenance Tagging
Evaluation Results
| Method | Links | |
|---|---|---|
| HMAC Provenance TaggingAttack class=Unsigned, Variant=direct, t=12026.06 | 0 | |
| HMAC Provenance TaggingAttack class=Unsigned, Variant=direct, t=32026.06 | 0 | |
| HMAC Provenance TaggingAttack class=Unsigned, Variant=flooding, t=12026.06 | 0 | |
| HMAC Provenance TaggingAttack class=Unsigned, Variant=flooding, t=32026.06 | 0 | |
| HMAC Provenance TaggingAttack class=Unsigned, Variant=subtle, t=12026.06 | 0 | |
| HMAC Provenance TaggingAttack class=Unsigned, Variant=subtle, t=32026.06 | 0 | |
| HMAC Provenance TaggingAttack class=Bypass, Variant=subtle, t=32026.06 | 0 | |
| SMSRAttack class=Unsigned, Variant=direct, t=12026.06 | 0 | |
| SMSRAttack class=Unsigned, Variant=direct, t=32026.06 | 0 | |
| SMSRAttack class=Unsigned, Variant=flooding, t=12026.06 | 0 | |
| SMSRAttack class=Unsigned, Variant=flooding, t=32026.06 | 0 | |
| SMSRAttack class=Unsigned, Variant=subtle, t=12026.06 | 0 | |
| SMSRAttack class=Unsigned, Variant=subtle, t=32026.06 | 0 | |
| SMSRAttack class=Bypass, Variant=subtle, t=32026.06 | 0 | |
| SMSRAttack class=Authenticated, Variant=direct, t=12026.06 | 37.8 | |
| SMSRAttack class=Authenticated, Variant=flooding, t=12026.06 | 43.1 | |
| HeuristicAttack class=Unsigned, Variant=direct, t=12026.06 | 86.7 | |
| HeuristicAttack class=Unsigned, Variant=direct, t=32026.06 | 86.7 | |
| HeuristicAttack class=Unsigned, Variant=flooding, t=12026.06 | 86.7 | |
| SMSRAttack class=Authenticated, Variant=direct, t=32026.06 | 92.7 | |
| UndefendedAttack class=Unsigned, Variant=direct, t=12026.06 | 93.3 | |
| UndefendedAttack class=Unsigned, Variant=direct, t=32026.06 | 93.3 | |
| UndefendedAttack class=Unsigned, Variant=flooding, t=12026.06 | 93.3 | |
| UndefendedAttack class=Authenticated, Variant=direct, t=12026.06 | 93.3 | |
| UndefendedAttack class=Authenticated, Variant=direct, t=32026.06 | 93.3 | |
| HMAC Provenance TaggingAttack class=Authenticated, Variant=direct, t=32026.06 | 93.3 | |
| HMAC Provenance TaggingAttack class=Authenticated, Variant=flooding, t=12026.06 | 93.3 | |
| SMSRAttack class=Authenticated, Variant=flooding, t=32026.06 | 96.2 | |
| UndefendedAttack class=Unsigned, Variant=flooding, t=32026.06 | 100 | |
| UndefendedAttack class=Unsigned, Variant=subtle, t=12026.06 | 100 | |
| UndefendedAttack class=Unsigned, Variant=subtle, t=32026.06 | 100 | |
| UndefendedAttack class=Bypass, Variant=subtle, t=32026.06 | 100 | |
| UndefendedAttack class=Authenticated, Variant=flooding, t=12026.06 | 100 | |
| UndefendedAttack class=Authenticated, Variant=flooding, t=32026.06 | 100 | |
| HeuristicAttack class=Unsigned, Variant=flooding, t=32026.06 | 100 | |
| HeuristicAttack class=Unsigned, Variant=subtle, t=12026.06 | 100 | |
| HeuristicAttack class=Unsigned, Variant=subtle, t=32026.06 | 100 | |
| HeuristicAttack class=Bypass, Variant=subtle, t=32026.06 | 100 | |
| HMAC Provenance TaggingAttack class=Authenticated, Variant=direct, t=12026.06 | 100 | |
| HMAC Provenance TaggingAttack class=Authenticated, Variant=flooding, t=32026.06 | 100 |