ResearchFlashAttention-4 1613 TFLOPs Blackwell 71 Percent UtilizationarXiv·high signalNew attention kernel for Blackwell GPUs. 1.3x over cuDNN, 2.7x over Triton. Written in CuTe-DSL Python. 4x perf at long sequences.SourceSource pagearXiv↳ Follow the threadStack layer / Update threadBoogu-Image-0.1 Tops HuggingFace Daily With 727 Upvotes: Frontier-Adjacent Image Generation Trained for ~$400KarXiv / HuggingFace Daily PapersStack layer / Threat patternAutomated Template-Free Synthesis of Instruction-Centric Leakage Contracts for Black-Box CPUsarXivPolicy dependency / Stack layerVerification is now harder than generation for coding agents — budget accordinglyarXiv 2606.26300Stack layer / Threat patternPlausible Deniability Guarantees for Whistleblowers: A Formal Privacy Framework for Anonymous ReportingarXivFollow-up threadSciDiagramEdit learns to edit scientific figures from arXiv revision histories via agentic skill evolutionarXivStack layer / ContrastOpenAI unveils GPT-Red, an internal LLM 'super-hacker' used to harden its modelsMIT Technology ReviewStack layer / Threat patternarXiv Paper Argues Agent Security Is a Systems Problem, Not a Model ProblemarXiv / Hacker NewsFollow-up threadNavigating the Socio-Technical Complexity Challenge in Quantum Software EcosystemsarXiv