Enkstein is a self-hosted MIT-licensed workspace that adds policy-aware routing, approval gates, and audit trails across Codex, Claude, Gemini, and Ollama agents.
doma is a single-binary BM25 search tool in Odin built to stop Claude Code from grepping wildly and missing relevant results across your codebase.
OpenAI's GPT-5.6 family — Luna ($1/$6), Terra ($2.50/$15), Sol ($5/$30) — hits GA. Sol beats Claude Fable 5 on Agents' Last Exam by 13.1 points.
Jon Udell argues developers should stop framing agentic coding as a machine-owned loop. The human stays in charge; agents join the team.
Simon Willison used Claude Code to convert the Moebius 0.2B inpainting model from PyTorch/CUDA to ONNX, running it entirely in-browser via WebGPU. The demo is live.
New research from MIT shows LLMs assign roles based on writing style, not role tags. Prompt injection works because untrusted text can simply sound like a system prompt.
Anthropic drops its policy of secretly degrading Claude Fable responses to frontier AI researchers after widespread backlash, promises visible safeguards instead.
Daniel Stenberg says AI-assisted security reports hit curl at 4-5x the 2024 rate. The quality is higher, the volume is unsustainable, and his wife is worried.
Simon Willison's datasette-agent 0.2a0 lets tools ask users yes/no, multiple-choice, or free-text questions mid-execution, with conversations that survive server restarts.
Anthropic published a detailed engineering post on how it sandboxes Claude across Claude.ai, Claude Code, and Cowork — with gVisor, Seatbelt, Bubblewrap, and full VMs.
Simon Willison spent 5.5 hours stress-testing Anthropic's Claude Fable 5. It's a big model with big guardrails — and a new fallback mechanism when safety filters trigger.
Anthropic's Claude Opus 4.8 arrives with improved honesty, better uncertainty flagging, and benchmark leads over GPT-5.5 — and the company actually calls it a minor update.
Developer David Wilson catalogued 16+ unfinished AI-assisted projects and concluded the fix might be cancelling his subscription entirely.
Apple's rebuilt Siri runs a custom Gemini-derived model on Private Cloud Compute and uses vision LLMs to read your screen — plus a new Core AI library for on-device model development.
Andreas Kling shut down public PRs for the Ladybird browser. The reason: AI-generated patches destroyed the assumption that big contributions signal good faith.