OpenAI’s Oct 4 prerelease keeps command/patch tools advertised when execution environments are starting or failed (model told to wait), tightens MCP approval routing so unrelated threads can’t surface prompts, and hardens Windows socket ACLs plus managed-daemon install retries. Pin @openai/[email protected]; npm latest is still 0.160.0.
OpenAI’s Oct 4 prerelease routes eligible remote control launches through the managed daemon, falls back to mklink when Windows daemon junction updates are denied, preserves review lifecycle order on failure, and honors server reasoning summary defaults in new TUI threads. Pin @openai/[email protected] alpha.13 ; npm latest is still 0.160.0…
NVIDIA’s Apache 2.0 SkillSpector checks Claude Code, Codex, and Gemini skills (plus MCP) for prompt injection, exfil, supply chain, and tool poisoning risks—71 patterns across 17 categories, with optional LLM analysis and an MCP scan skill install gate. A 31k skill study found 26.1% vulnerable; install via uv tool install 'skillspector[mcp] @ git+https://github.com/NVIDIA/skillspector.git'…
llama.cpp’s server now exposes POST /v1/systemone (System One / Jev compatible): send state plus choice, score, or noul questions and get typed probabilities in one forward pass. Ships with Julia 1, Laya, Kev 4B, lev, and image capable OpenJev—start with llama serve hf ggml org/Kev 4B GGUF…
Baseten used Claude Code (Fable 5) plus MetaInfer skills to vibe build VibeQwen—a Qwen 3.6 35B A3B NVFP4 engine on one B200 that beat tuned vLLM 0.25.1 by up to 90% single stream decode (1,792 vs 943 TPS) and cut TTFT 28→12ms; a follow on SAM 3.1 server gained 50% throughput. Experimental only—not serving production yet…
Microsoft and Hugging Face released ThinkingBox: 507 stateful business workflows graded on terminal backend state and side effects across 20 repeats—not tool call trails or final replies. Run it via OpenEnv; Claude Opus 5.5 leads pass@1 ( 67%), and 80% of failures are tool handling, not reasoning…
Spotted something we missed? Start a thread.
