ResearchMOSAIC Safe Multi-Step Tool Use Plan-Check-Act FrameworkarXiv·high signalXBlueskyLinkedInCopy linkPost-training framework for safe agent tool use, 50% harmful behavior reduction, 20%+ refusal on injectionSourceSource pagearXiv↳ Follow the threadPolicy dependency / Stack layerSCA-Agent reconstructs a dependency's trace across Code, Build, Release, Deploy and Runtime and beats the best traditional SCA tool by 18.76 points of F1arXiv 2609.18391Stack layer / Threat patternPentestChain keeps a 7B local model off the critical path behind a deterministic exploit map and runs an eleven-tool MCP pentest pipeline at zero paid-API costarXiv 2609.18120Stack layer / Threat patternA Model Can Fingerprint vLLM or SGLang From Its Own Output Tokens, Then Exploit ItarXiv 2609.20614Stack layer / Threat patternA Fake Security-Product Story in an Unexecuted Binary Section Flipped 30 of 35 Malware VerdictsarXiv 2609.19722Stack layer / ContrastModel-Checking an Agent's Plan Before Any Tool Runs Rejects Unsafe Plans Without Spending a Single Tool CallarXiv 2609.18674Stack layer / ContrastAdversarial Agents Ran Arbitrary Bash Past Claude Code Auto Mode and Codex Guardian in 79% of TrialsarXiv 2609.19587Stack layer / Update threadRaising a Stated Failure Probability From 10% to 70% Changes Whether Frontier Models Check the Evidence by at Most 21 PointsarXiv 2609.17865Policy dependency / ContrastThree SBOM Generators Diverge Systematically on 3,000 Projects, 14 Months Before the CRA Makes SBOMs MandatoryarXiv 2609.19920