LLM Agent Security Evaluation on AgentDojo vulnerable-agent profile stress test (six banking pairs)
0Attack Success Count (out of 6)Deny
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DenyDefence=Deny, System=MemLineage2026.05 | 0 | 2 | 6 | 0 | |
| Strip+retryDefence=Strip+retry, System=MemLineage2026.05 | 0 | 4 | 31 | 0 | |
| Authority repairDefence=Authority repair, System=MemLineage2026.05 | 0 | 4 | 4 | 4 | |
| No defenceDefence=No defence2026.05 | 6 | 2 | 0 | 0 | |
| Sig. onlyDefence=Sig. only2026.05 | 6 | 2 | 0 | 0 |