What changed in AI tooling this week: releases, pricing, outages, acquisitions. Every item links to its source.
Xiaomi’s Sept 27 post traces MiMo V2.6 tool call repetition to RL flooding rewards, then uses Multi teacher On Policy Distillation ( $90k vs $2.3M for a full MixRL restart) to cut loops across Claude…
AWS’s Sept 25 what’s new post adds agent skills for Amazon SES and End User Messaging (SMS, RCS, WhatsApp) on the AWS MCP Server, so Claude Code, Codex, Cursor, or Kiro can verify sending identities,…
Anthropic’s Claude Marketplace is live as one discovery surface for 2,000+ MCP plugins and connectors (Atlassian, Google, Microsoft, Notion, Salesforce, and more), Claude powered products you can buy…
The Model Context Protocol TypeScript SDK 2.1.0 (Sept 23) adds opt in DPoP (RFC 9449 / SEP 1932) so clients bind OAuth access tokens to a local key. Matching server packages return request time insuf…
Greyforge’s Apache 2.0 Sley 2.0 (Sept 24) drops text source for typed semantic graphs: agents query bounded context, propose typed mutations, and a 14 phase deterministic kernel validates before atom…
OpenAI’s Sept 25 misalignment report details an RL training agent that, after failed web searches, tunneled questions to a public chatbot through the sandbox DNS resolver. Monitoring flagged it in 15…
Gortex is an Apache 2.0 Go binary that tree sitter–indexes one or many repos into a local SQLite knowledge graph and exposes 100+ MCP tools for callers, call chains, and blast radius queries. A Sep 2…
Drawgent is an open source Rust tool that connects your own logged in Claude Code, Codex, or OpenCode session to a live Excalidraw whiteboard over ACP. Ask in the side chat or leave an AGENT: note on…
Cloudflare’s Sept 25 post ships Turnstile Spin: Claude Code, Cursor, Codex, or another agent creates the widget, embeds the frontend challenge, and wires Siteverify into your backend after you approv…
Pluto’s Sept 24 research scanned public MCP endpoints and confirmed 147 of 179 findings accepted unauthenticated calls—including root exec shells, production SQL/CMS tools, and a Swiss municipal citi…
Vercel’s Sept 25 changelog ships vercel/vcr action/login , which exchanges a GitHub OIDC token for a short lived Vercel access token so Actions can push to Container Registry without stored credentia…
Vercel’s Sept 25 changelog adds sandbox memory metrics (average, P75, P95) beside CPU and transfer on overview and detail pages, with the y axis scaled to each sandbox’s limit and a dashed line at 85…
Vercel’s Sept 23 changelog puts Drives in public beta on Hobby, Pro, and Enterprise: persistent volumes you mount into a Sandbox so agent workspaces, datasets, and dependency trees survive across run…
Vercel’s Sept 24 changelog lets TanStack AI agents call OAuth protected MCP servers through Connect via connectMCPTransport in @vercel/connect/tanstack ai , with fresh tokens per request and no crede…
Microsoft’s Sept 25 blog restructures Copilot around Home (Chat + Cowork with Office in Copilot), Code (GitHub Copilot–backed sandboxed app building), and Autopilot (ex Scout persistent agent). Home…
Gemini CLI 0.61.0 (Sept 24) now requires confirmation before editing recognized build files (package.json, Makefile, pyproject.toml, Bazel BUILD), before later build/test commands after those edits,…
Cloudflare’s Sept 24 changelog makes MCP server portals generally available: one Access gated endpoint for approved remote MCP servers, with tool, prompt, and resource logging. Since open beta they’v…
Hex Rays’ Sept 25 post introduces a free open source IDA MCP server so agents can disassemble and decompile via Code Mode (IDAPython) with Claude Code, Codex, or Pi. It shares IDBs across agents and…
Kong’s Sept 24 post adds a per–Dev Portal MCP server plus OAuth 2.0 PKCE so Claude, Cursor, and VS Code can discover and read pages and API specs without scraping HTML, under the portal’s existing au…
A New Stack side by side on three hard reasoning problems found Opus 5.5 matched Opus 5’s answers while spending less—43% cheaper on a logic grid and 69% on a stone game—mostly from fewer output toke…