cs.CRNov 14, 2026

VAST: Compiling Grounded Vulnerability Scenarios into Verifiable, Multi-Step Environments for Evaluating LLM Red-Team Agents

Authors: Zhen Wan

Organizations: SCANTIST PTE. LTD.

Abstract

Evaluating Large Language Model (LLM) security agents requires reproducible environments whose workflow difficulty can be varied and whose failures can be diagnosed. We present VAST (Verifiable Adversarial Scenario Training & Evaluation), a declarative range compiler separating vulnerability scenarios from a compiler-introduced workflow-hardening layer. Following a post-hoc code and AST audit across all topics, the main benchmark retains 109 seed edges (45 with strict upstream CVE annotations), 72 family identifiers (62 under shared-HMAC relay grouping), 81 topic configurations (76 core plus 5 bounded execution models), and 1,387 instances (359 canonical four-gate-revalidated specs plus 1,028 seed variants) in a 42:39 partition. On the TopicServer compiler stratum, a budget-matched 270-episode prompt contrast across 10 compiled mechanism-family clusters (30 P2/P3/P4 instances, 40 turns, temperature 0.0) shows that glm-5.3, grok-4.7, and deepseek-v4-pro solve 12/90 instances under contract_only and 14/90 under generic_hint (cluster sign p=1.00), whereas supplying the shared workflow protocol in full_guide lifts completion to 76/90 (28/30, 23/30, 25/30; cluster sign p<=0.00391). On the separate ranges_v2 ledger, a 126-episode evaluation across 6 non-TopicServer workflows and multi-container Docker/Linux stacks solves 88/126 (69.8%) (28/42, 31/42, 29/42), while an 18-episode intervention on asynchronous Redis worker chaining (v2_02) lifts solves from 0/18 to 18/18 (100%) in 4-6 turns when installed worker modules are disclosed in GET /sched/directory. Finally, a three-seed (s in {42, 1024, 2025}) LoRA recipe comparison on Qwen2.5-Coder-7B-Instruct across 37 lineage-disjoint families (231 episodes on 33 held-out instances) yields 0/33 for Base, 1.67 +/- 2.89 / 33 for Golden SFT, and 4.67 +/- 2.31 / 33 for Header-Recovery SFT, with all solves on v2_03.

Figures & tables

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Jul 13, 2026

Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming

Production LLM agents such as Claude Code and Codex operate over untrusted content, files, commands, and workspace state, making safety failures directly actionable. Red-teaming must therefore keep pace with evolving models and tools. Existing approaches mainly optimize attack success and preserve artifacts such as benchmarks, payloads, or attack programs, which record where attacks succeed but not the enabling conditions behind unsafe agent behavior. We study automated red-teaming for production LLM agents using one agentic research environment to discover reusable vulnerability knowledge about another. We present AHA, a falsifiable discovery loop that proposes a vulnerability hypothesis, constructs a falsifier, instantiates a valid attack, executes it in a sandboxed harness, reflects on the trajectory, and promotes confirmed findings into a Vulnerability Concept Graph (VCG). Each concept links an attacker-facing surface to an unsafe trajectory through a claim, enabling condition, falsifier, transfer prediction, and supporting evidence. Across Claude Code and Codex on three scenarios covering direct and indirect attacks, the discovered concepts reveal a reusable vulnerability core across models and agents. A frozen VCG requires no further search and outperforms the strongest frozen discovery baseline by 14.2 percentage points under the same single-shot protocol, while transferring across scenarios and attack channels. The resulting VCG provides an auditable artifact for production safety teams to inspect vulnerabilities, validate patches, and accumulate reusable safety knowledge. Our code is available at https://github.com/henrymao2004/Auto-research-red-teaming-in-sleep.
Jun 7, 2026

ForesightSafety-SAGE:A Fully Automated Scenario Generation and Safety Evaluation Framework for LLM Agents

Large language models (LLMs) are increasingly evolving from simple text-based interaction systems into LLM agents that can maintain memory, use tools, access external environments, and execute tasks. As their capabilities and autonomy expand, the safety risks they face also become more diverse. Existing evaluations often rely on manually written scenarios, static prompts, or final-output judgments, making it difficult to capture the diverse risks that agents may face during task execution. We introduce ForesightSafety-SAGE, a fully automated scenario generation and safety evaluation framework for LLM agents. Based on five risk dimensions,we instantiae abstract and diverse safety risks in real-world task execution into 1,072 measurable evaluation scenarios. Using the automated evaluation pipeline, 12 LLM agents are evaluated under two authority contexts. The results show that current agents still face substantial behavioral safety risks during task execution, with an average ASR of 47.1% and several models exceeding 70%. These findings demonstrate the importance of executable, process-level evaluation for understanding and improving LLM agent safety.
Oct 5, 2026

Correct Verdicts, Flawed Reasoning: Structured Auditing of LLM-based Vulnerability Reasoning

Large Language Models (LLMs) are increasingly deployed for automated software vulnerability analysis. Binary classification alone is insufficient; practitioners need explanations to triage bugs and engineer patches. Standard practice relies on Chain-of-Thought (CoT) prompting, but free-form reasoning allows models to obscure logical leaps, hallucinated execution steps, and internal inconsistencies behind plausible prose. Our manual audit reveals that approximately 60% of correct vulnerability verdicts are accompanied by fabricated or unverifiable claims, and free-form explanations allow reasoning errors to evade LLM-as-a-judge evaluation. We present Vulnerability Explanation Reasoning Auditor (VERA), an automated framework for auditing LLM vulnerability reasoning. Rather than accepting free-form text, VERA asks models to output a Structured Reasoning Record (SRR) encoding tracked pointers, memory operations, and state transitions in machine-readable fields. A multi-stage judge audits each SRR against eight reasoning failure modes using deterministic checks, with LLM calls reserved for semantic interpretation. The standardized SRR schema also enables automated mutation testing to benchmark judges at scale without human annotation. Our evaluation shows reasoning flaws occur in correct verdicts just as frequently as incorrect ones, and VERA exposes 87% of reasoning errors that free-form LLM-as-judge systematically miss.