NUWA Evidence record
Qwen2-72B
Agent Security Bench (ASB) · Evidence record
Original source charts and materials
Recorded measurements
- Attack success rate
- 53.7%
Not disclosed
ASB ReAct framework based on AIOS · Budget not specified
Method and scope
Which agent interfaces expose distinct attack and defense failure modes?
- Evidence type
- Benchmark
- Assessment subject
- ASB separates system/user prompts, tool responses and retrieved memory, then evaluates attacks and defenses on a shared agent framework.
- Primary dimension
- Behavior
- Evidence snapshot
- 2026-10-29.1
Attack surfaces, backbones, and defense configurations differ; its aggregate attack rate is not a deployment probability. Revision-dependent method and metric counts differ from older abstracts.
Source finding
The captured revision reports vulnerability across multiple interfaces and limited protection from the evaluated defenses. Utility and security must be assessed together.
Limitations
Attack surfaces, backbones, and defense configurations differ; its aggregate attack rate is not a deployment probability. Revision-dependent method and metric counts differ from older abstracts.
Attack surfaces, backbones, and defense configurations differ; its aggregate attack rate is not a deployment probability. Revision-dependent method and metric counts differ from older abstracts.
Related research questions
References
Numerical results are reported by the cited authors.
- [1] Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents. 2024-10-03 ; #S5.T5; row Qwen2-72B; condition Average across attack types; column Attack success rate