ResearchAgentic Critical Training — Why Actions Succeed Not Just What To DoarXiv·high signalNew training paradigm contrasting successful against suboptimal actions. Quality judgment as first-class training objective.SourceSource pagearXiv↳ Follow the threadStack layer / Threat patternMako: A 'Self-Evolving Agentic OS' That Writes New Exploits Against Live Targets and Hot-Loads Them Into ItselfarXivStack layer / ContrastBackendForge: Best Coding Agent (GPT-5.5) Deploys Correct Backends Only 28.6% of the Time Under a Strict OraclearXivPolicy dependency / ContrastNetInjectBench: Naive Network-Ops Agents Take Unsafe Actions 82.5% of the Time Under Prompt InjectionarXivStack layer / ContrastStore agent episodes in a dedicated event subgraph and use cross-graph traversal for multi-hop temporal questionsarXiv 2604.12285Stack layer / Threat patternMCPZoo Study: Scanners Call 96.89% of MCP Servers 'Risky' — But Under Half of Those Alerts Are RealarXivContrastEvidence-Backed Video Question Answering pushes Video LLMs from black-box answers to grounded evidencearXivContrastExplainable Agentic Detector Flags Conversational Scams at 97.8% AccuracyarXivStack layer / Threat patternStudy of 44 Developers: AI Code Assistants Barely Improve Secure API UsagearXiv