ResearchDefensive Refusal Bias Safety Alignment Fails Cyber Defenders 2.72xarXiv·high signalSafety-tuned LLMs refuse legitimate defensive security tasks at 2.72x rate, system hardening 43.8% refusalSourceSource pagearXiv↳ Follow the threadPolicy dependency / Stack layerReason Less, Verify More: Deterministic Gates Catch a Silent Policy-Violation Failure in Tool-Using AgentsarXivStack layer / Threat patternSLBench Finds Up to 70% Unsafe Execution When Codex and Claude Code Follow Logical Relations in Skill FilesarXivThreat pattern / Follow-up threadVEXAIoT: arXiv paper demonstrates autonomous IoT vulnerability exploitation using AI agentsarXivThreat patternDual-Graph Verification Framework Audits Legacy IT Security Concepts for NIS-2 / OSCAL MigrationarXivStack layer / Threat patternMitigating Taint-Style MCP Server Vulnerabilities Without Touching Server Code (SPELLSMITH)arXivStack layer / Threat patternLong-Horizon-Terminal-Bench: Best Frontier Agent Clears Just 15% of Real Terminal Tasks, Mean 4.3%arXiv (Tencent Hunyuan) / HuggingFace Daily PapersThreat patternCAAD Reframes Time-Series Anomaly Detection as Granger-Causality Consistency CheckingarXivStack layer / Threat pattern'Resample or Reroute?' Reframes LLM Routing Under Real Budget ConstraintsarXiv