Benchmark RadarAgentS4D
AgentS4D evaluates runtime safety of LLM-based workspace agents across a four-dimensional framework, with 328 risk-injected cases and seven lifecycle checkpoints, measuring unsafe behavior and evidence across six risk-entry sources and nine harms.
- Released
- 2026-07-29
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing safety benchmarks focus on endpoints, missing risks that emerge during execution. This benchmark provides a structured way to assess agent safety across the lifecycle, showing that task completion does not imply safety and that testing one risk form can miss vulnerabilities.
Motivation
Large language model (LLM)-based workspace agents execute stateful, multi-step workflows across heterogeneous resources, external tools, and persistent state.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.