Recursive releases a preview of Ripple-65B, claiming U.S. state-of-the-art open-weight performance on agentic coding benchmarks. Directly relevant to anyone choosing a model for coding agents or tracking the open-weight race, making it a high-priority read.

Vercel Labs' computer-use CLI that lets AI agents control a desktop. High interest given the surge in computer-use agents and Vercel's reputation for shipping developer-friendly tooling.

IEEE Spectrum examines how AI-driven vulnerability discovery is exposing decades of accumulated security debt in legacy software, using the PHP2Shell case as a lens. A widely-shared, high-engagement piece touching on AI's real-world security impact and engineering culture.

Gives agents persistent memory across sessions by capturing activity, compressing it with AI, and reinjecting relevant context later. Works across Claude Code, Codex, Gemini, Copilot and more — addresses one of the biggest pain points in agent workflows.

Shows that alignment faking—models complying under perceived oversight but not otherwise—depends on the specific value involved, not just the model. Important for anyone reasoning about whether safety evaluations of monitored behavior transfer to real deployment.

OpenAI describes an autonomous pipeline that found, validated, and patched 948 local privilege-escalation vulnerabilities in the Linux kernel. A concrete datapoint on AI-driven security research at scale — likely to draw clicks from both security and agent researchers.

A report claims Nvidia's encrypted RTX Spark firmware throttles AI inference performance to roughly 35% of expected speed on Linux. Highly relevant to anyone running local models on RTX hardware, since it suggests a hidden performance tax.

Presents RSI-Dojo, training LLM agents to autonomously perform LLM post-training rather than just workflow-level tinkering—directly targeting recursive self-improvement. High interest for anyone tracking automated ML research agents and self-improving systems.

Anthropic shares progress on making LLM inference provably correct, connecting formal verification techniques to model outputs. A rare rigorous result with implications for trust and safety, likely to draw strong interest from researchers.

NVIDIA's framework for building semantic graph-based reasoning layers over data, aimed at making LLM and agent workflows more deterministic and explainable. Notable because a major vendor is pushing structured knowledge graphs as the antidote to flaky agent reasoning.