Strick Sheng

Ziyue Wang, Cheuk Wang Maurice Ng, Chenchen Yu et al.

An LLM agent for vulnerability discovery and validation is more than a model. It combines three components: an LLM for code analysis, an agent harness such as Codex or OpenCode for navigation, tool use, and execution, and an audit playbook, domain-specific procedural knowledge that guides the LLM and harness toward vulnerability discovery. Prior work relies on human-supplied playbooks, including prompt engineering, manual workflows, knowledge bases, and heuristics. This raises two research questions: Acquisition - is human curation necessary, and can playbook creation be automated? Transfer - can an evolved playbook transfer the audit procedure to weaker agents, improving their capability? We present EvoHunt, a playbook evolution environment over open-source repositories for security auditing. Three agents drive the evolution loop: an audit agent rolls out the current playbook and produces findings; an evaluator scores outcomes against ground truth; and a reviser commits updates to the playbook based on failure analysis. The playbook format is unconstrained: starting empty, EvoHunt adds or removes workflows, heuristics, vulnerability knowledge, or domain-specific content. The evolved playbook requires only minor adaptation to run under a different LLM or harness. We evaluate EvoHunt on open-source security advisories. For acquisition, playbook evolution raises end-to-end exploits for Codex/GPT5.4-xhigh 6x, from 1.1% to 6.2%, and the evolved OpenCode/GLM5.1 playbook surpasses OpenAI Codex Security on every metric, with 11.3% vs. 9.2% target-match rate, showing open-source evolution can outperform a dedicated commercial product. For transfer, the GLM-evolved playbook gives the strongest student lift: Qwen3.6-27B improves from 2.4% to 6.5%, Qwen3.6-35B-A3B from 1.1% to 4.6%, and A3B obtains 2.4x more matches than GPT transfer.

Strick Sheng, Ziyue Wang, Liyi Zhou

Large language model agents increasingly operate through environment-facing scaffolds that expose files, web pages, APIs, and logs. These observations influence tool use, state tracking, and action sequencing, yet their reliability and authority are often uncertain. Environmental grounding is therefore a systems-level problem involving context admission, evidence provenance, freshness checking, verification policy, action gating, and model reasoning. Existing agent benchmarks mainly evaluate task capability or specific attacks such as prompt injection and memory poisoning, but they under-specify a fundamental reliability question: whether agents remain grounded in the true environment state when observations are stale, incorrect, or malicious. We introduce EnvTrustBench, an agentic framework for benchmarking this failure mode. We define an evidence-grounding defect (EGD) as a behavioral failure in which an agent treats an environment-facing claim as sufficient evidence for action without resolving it against available current evidence, leading to a task-incorrect false path under the true environment state. Given a task scenario, EnvTrustBench generates the workspace, environment, agent-facing objective, and validation oracle, executes the evaluated agent, records its action-observation trajectory and final state, and applies the oracle to produce a verdict. Using 6 LLM backbones and 5 widely used scaffolds, we evaluate 55 generated cases across 11 task scenarios, with each scenario expanded through five feedback-guided generation iterations. Results show that EGDs consistently emerge across operational workflows, highlighting environmental grounding as a core agent reliability problem with important security implications.