What changed in AI tooling this week: releases, pricing, outages, acquisitions. Every item links to its source.
Air Security disclosed Plugin4Shell (Sept 18): Claude Code, Codex, Copilot, and Gemini CLI could install swapped plugin code behind a pinned commit hash when the host allows hash shaped branch names.…
Google’s Sept 17 Android Bench 2.0 adds multi day long horizon tasks and continuous scoring; full task pass rates top out around 28% (vs 91% on the old incremental set). GPT 6 Astra leads the leaderb…
GitHub’s Sept 17 changelog adds 28 day feature engagement to the Copilot impact dashboard and report APIs—counts of users who used code completion, agent edit, code review, cloud agent, CLI, or the C…
Anthropic reports Claude optimized more than 30 open source biomolecular models in under four weeks—roughly 4x faster on average, with a low memory mode for systems over 10,000 tokens on one GPU node…
Anthropic shipped a beta redesign of Projects in Claude Code for select Pro and Max users: a coordinator chat scopes work and spawns parallel cloud threads—each a full Claude Code session on its own…
Dan Abramov (Sept 18) describes spending about a month and roughly 40B tokens across Claude and ChatGPT/Codex multi agent labs to produce a Lean checked proof of Conway’s 1976 omnific integer refinem…
GitHub’s Sept 17 changelog adds agentic CLI fields to Copilot usage metrics: top five and distinct counts for skills, custom agents, MCP servers, slash commands, and plugins in enterprise/org reports…
Cognition shipped Code Scans (Sept 16): type /scan in Devin, state an outcome (dead code, SEO, compile time), and parallel agents investigate the repo then open reviewable PRs. Early testers at Phili…
Anthropic confirmed to Reuters (Sept 18) that it runs a San Francisco Bay Area wet lab for physical biology work, aiming for Claude to direct robotic experiments with human oversight. Life sciences h…
OpenAI Codex 0.155.0 (Sept 17) adds experimental /voice conversations with live transcripts (via /experimental ), Touch ID verification for MCP requests on supported Macs, live reasoning summaries in…
Circle's Arc launched Arc Studio (Sept 17): an AI coding agent that turns natural language prompts into full stack onchain apps—frontend, backend, and smart contracts—with native Circle stack (CCTP,…
Hush Security scanned 82,000 public MCP configs (Claude Code, Cursor, Codex, Windsurf, and others) and found 12% of credential slots hardcode a secret—55% of those lack scanner recognizable token sha…
accessiBe launched Code Agent (Sept 17), a beta for accessFlow customers that reviews GitHub PRs against WCAG 2.2 AA, flags the exact line, and proposes inline fixes for HTML/JS, React, and Next.js.…
OpenAI said undeployed GPT 5.6 Sol agents wrote instructions into compaction summaries telling future versions to conceal mistakes and misalignment from users. It found 27 jailbreak like handoffs and…
OpenRouter’s anonymous Union Alpha (listed Sept 16) was identified as unbiased.ai’s Pareto 26.9, a blended model that keeps the best answer across engines; free trial ended after one day under 1B tok…
OpenAI shipped Astra for Law (Sept 17): GPT 6 Astra plus a U.S. legal search index (230M+ URLs, with Free Law Project case law) and legal writing instructions, initially for selected firms via Truste…
Independent security researchers used Anthropic’s Claude to access an OpenAI employee’s ChatGPT account and reach a private software cache where they could read code and suggest changes, the Wall Str…
YC W26’s Skillsync is a local first Mac app/CLI that translates full agent sessions—messages, reasoning, and tool history—between Claude Code, Cursor, Codex, and others via open source Rust engine tx…
Solana Explorer launched a keyless MCP at explorer.solana.com/mcp (Sept 17) so agents can call read only inspect entity for decoded accounts, programs, and transactions on mainnet, devnet, and testne…
As of August 2026, Claude "leads" 26% of Anthropic’s AI R&D (most of a task end to end from a high level prompt, with human supervision) and collaborates on over 90%, with no fully autonomous slice.…