Part of the benchmark canon: 1 published, 1 queued next, 6 not yet run. How to read the Paladin column is defined on the canon page.
| Benchmark | Organization | Measures | Paladin |
|---|---|---|---|
| InjecAgent (top 12) | UIUC Kang Lab | 1,054 indirect prompt-injection cases against tool-using agents | Results published |
| AgentDojo (top 12) | ETH Zurich / Invariant Labs | Utility and injection resilience for tool-using agents, scored jointly | Queued next |
| AgentHarm (top 12) | Gray Swan AI / UK AISI | Agent harmfulness when the user is the attacker | Not yet run |
| Agent Red Teaming (ART) | Gray Swan AI | High-impact attacks curated from a large public competition | Not yet run |
| SHADE-Arena | Anthropic | Sabotage and monitoring across 29 agent environments | Not yet run |
| BIPIA | Microsoft | Indirect prompt-injection attacks and defenses | Not yet run |
| Tensor Trust | UC Berkeley | Prompt extraction and hijacking, drawn from a public game | Not yet run |
| BrowserART | Gray Swan AI | Red-teaming for browser agents | Not yet run |
Next: Cyber offense and CTF.