This work introduces AgentS4D, a sandboxed benchmark for lifecycle-wide runtime safety evaluation and evaluates all 20 combinations of four harnesses and five LLM backends, finding that the observed safety of an agent system varies with both its harness-LLM pairing and how risk is introduced.
Abstract
Large language model (LLM)-based workspace agents execute stateful, multi-step workflows across heterogeneous resources, external tools, and persistent state. Their safety must therefore be assessed from actions, side effects, and state changes throughout execution. Although recent benchmarks have advanced executable safety testing and trajectory-aware verification, they rarely provide a unified account of where risks enter, how they elicit unsafe behavior, which harms they target, and where supporting evidence appears during execution. We introduce AgentS4D, a sandboxed benchmark for lifecycle-wide runtime safety evaluation. Its four-dimensional runtime-safety framework uses six risk-entry sources, six induction strategies, and nine target harms to guide case construction, while seven lifecycle checkpoints organize post-run evidence. AgentS4D contains 328 risk-injected cases. We evaluate all 20 combinations of four harnesses (Hermes, OpenClaw, Claude Code, and Codex) and five LLM backends (GPT-5.5, Gemini 3.1 Pro, DeepSeek-V4-Pro, MiniMax-M3, and Qwen3.7-Plus) on these cases, yielding 6,560 runs. Overall, 4,461 runs (68.0%) trigger prespecified unsafe signals. Across the 20 configurations, the observed safety of an agent system varies with both its harness-LLM pairing and how risk is introduced. Agent systems exhibit markedly different safety behavior when the same induction strategy reaches them through different risk carriers. They also respond differently to the same target harm when it is realized through different carriers and strategies. Moreover, 4,344 runs (66.22% overall) are unsafe yet complete. Thus, task completion cannot establish runtime safety, and testing only one form of a risk can conceal important weaknesses. Evaluations should examine complete agent configurations across diverse risk conditions and retain evidence throughout execution.
Dynamic agent harnesses let language models change the software that shapes their own execution. This flexibility brings a new reasoning burden: a local plugin change can propagate through dependencies and cleanup. We introduce CordisBench, a 1,200-question benchmark of this lifecycle reasoning. It combines a controlled formal setting with programs executed against Cordis, a runtime that manages component dependencies and cleanup, and asks models to identify affected components, predict state after a specified teardown order, determine which conditions hold under all or some orders, and choose reconfigurations that succeed when executed. Across these tasks, we evaluate three efficiency-oriented models at low reasoning effort with 2, 4, 8, 16, 24, or 32 relevant interactions, using deterministic task-specific scoring. Models usually handle small systems well but grow less reliable as more interactions become relevant, especially when predicting final state and when reasoning across teardown orders. Additional inference effort recovers marked gains for some models. The cost is nontrivial: on our 16-interaction subset, GPT-5.6 Luna uses nearly 3,000 reasoning tokens per question at medium effort. For these controlled instances, that cost is avoidable: an independent finite reference semantics agrees with Cordis execution on every observation and action outcome used for scoring across all 528 executable questions.
RedAgentBench is introduced, an executable framework for autonomous red-teaming and faithful measurement that shows that executable evaluation can improve safety measurement and identify actionable intervention points.
Zixing Chen, Xingyuan Liu, Jie Zhu et al.· 1 citation
This paper proposes SafeKeep, an inference-time safeguard that decouples safety judgment from tool execution: it assesses requests using flattened textual tool specifications while retaining the original schema-formatted specifications for execution.
OpenJiuwen provides a shared execution substrate and Rail-based capability composition across single agents, delegated sub-agents, and Swarm Flow, enabling developers to construct sophisticated agent harnesses under common execution semantics.
openJiuwen Team Tao Yu, Xin-Yu Zhang, Qian-Qian Chen et al.· 0 citations
AgentRewind is presented, a runtime recovery framework that records aligned checkpoints of the agent context and controlled environment, allowing agents to return to an earlier state and resume execution with information from previous attempts, improving task success rate and average checklist progress over the compared baselines.
Yu Zhuang, Kefei Chen, Yitong Duan et al.· 3 citations
Multi-agent systems built on large language models (LLMs) are increasingly deployed for complex tasks requiring autonomous planning, tool use, and inter-agent coordination. However, the non-deterministic nature of LLM outputs and the emergent behavior arising from agent interactions render traditional test oracles ineffective, creating a critical gap in quality assurance for agentic AI. This work introduces MORPHAGENT, a framework designed to address the oracle problem in multi-agent LLM systems through trace-based behavioral analysis. Our contributions are threefold: (1) goal-preservation relations that verify consistent goal achievement under input perturbations, (2) coordination-consistency relations that validate inter-agent delegation and communication patterns under agent substitution and reordering, and (3) tool-use integrity relations that ensure semantic equivalence of tool invocation sequences under prompt paraphrasing. MorphAgent instruments agent execution to capture structured traces comprising planning steps, tool calls, message exchanges, and final outputs, then systematically applies metamorphic transformations and checks behavioral invariants without requiring ground-truth oracles. We evaluate the framework on four multi-agent benchmarks spanning code generation, research synthesis, customer service, and data analysis tasks, encompassing 2,840 source-followup execution pairs across three LLM backends. Results show that MORPHAGENT detects 82.0% of seeded behavioral faults, including 90.3% of coordination failures and 81.7% of goal-deviation faults, while maintaining a false positive rate of 6.1%. The framework uncovers 14 previously unreported behavioral anomalies in established multi-agent frameworks, demonstrating its practical utility for assuring agentic AI reliability. These results suggest that trace-based metamorphic testing can serve as a practical foundation for reliable validation of emerging agentic AI systems.
Gopalakrishnan Marimuthu· International Conference on...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
