Trending AI Briefing: Tuesday, November 17, 2026 (morning ET)
The news this morning is mostly about files. Whoever can edit a rule file, a config file or a script decides what an agent does next, and Monday's posts pushed those files further from the work they judge. The Claude Code team explained that Code Review loads REVIEW.md from the base branch, so a pull request cannot rewrite the rules that review it, as long as that branch is protected. GitHub let enterprise admins block Copilot skills, hooks and MCP servers that come from loose user or project config, so they have to arrive as plugins or managed settings. Then the Optimizely incident showed the other side: a script from a trusted domain, rewritten by someone with write access, served by at least 178 organisations' sites in one day.
What's hottest in AI news right now
The Claude Code team published how it verifies agent-written code, and the sharpest rule is that a change can never edit its own check. The November 16 post on claude.dev, by Addy Osmani, says about 90% of the team's code is written and reviewed by Claude. Its Code Review feature reads REVIEW.md from the base or default branch, "never from the pull request's changes," and the team keeps every gating rule on a branch the change cannot touch. The docs add a condition: several of those rules only apply from a protected default branch, so "protect your default branch" is step one. A few other details are worth copying. A verify skill in .claude/skills/ boots and drives the app before every commit. Claude stops and calls a human if a check still fails after three rounds. A 39-item reliability checklist fires only the items whose trigger shows up in the diff, usually 3 to 10. A commit gate runs in off, observe and enforce modes, and it lets a third attempt through after two blocks so Claude does not get stuck. The catch is that this describes Anthropic's own repos. Most of it is a practice you build, and only a few pieces, such as Code Review's base-branch REVIEW.md, ship as product behavior. claude.dev · Code Review docs
GitHub let enterprise admins require that Copilot customizations come from plugins, on November 16. The new strictPluginOnlyCustomization managed setting blocks standalone skills, custom agents, hooks, MCP servers and executable Copilot extensions that come from user or project config. Admins can apply it to all five categories or pick some, for example locking down MCP servers while leaving skills open. GitHub is blunt about the limit: plugin-only "isn't a marketplace allowlist," so you pair it with strictKnownMarketplaces to control where those plugins come from. It is available on GitHub Enterprise Cloud for Copilot Business and Copilot Enterprise seats. GitHub Changelog
VS Code 1.146 deprecated the Local harness the same day, with removal set for February 2027. The Local harness is the older layer that runs Copilot agent sessions in the editor. Its replacement, the Microsoft Copilot harness, has been the default for new sessions since VS Code 1.143 and shares one runtime with Copilot CLI and the GitHub Copilot app. The governance angle is in one sentence: admins can enforce managed runtime permissions and sandbox restrictions that the Local harness cannot. The changelog adds that choosing the Copilot harness "doesn't enable sandboxing," so you configure that yourself. It also warns that switching harnesses doesn't convert every customization automatically. GitHub Changelog
Attackers wrote malware into Optimizely Web snippets on November 16, and sites including Zoom, the BBC and GitLab served it. Scott Helme's Report URI write-up says the attacker added code to each customer's Project JavaScript field, so the file still came from cdn.optimizely.com with a valid certificate. Report URI checked 1,611 snippets and found 188, belonging to 178 organisations, infected, and says the real number is likely higher. The payload swapped crypto wallet addresses, skimmed cards typed into a site's own form fields, stole non-HttpOnly cookies and screenshots, and loaded the Coruna and DarkSword exploit kits on older iPhones. The attacker's wallets had received about $7,040 by 19:25 UTC. How the attacker got write access across unrelated customer projects is still unknown, and Report URI says that is Optimizely's to explain. We did not find an Optimizely statement this morning, so every figure here is Report URI's. Report URI
Claude Code 2.1.326 reached npm at 18:43 UTC on November 16 with a run of fixes for rules that looked saved but weren't. The release adds a startup warning when an --allowedTools, --disallowedTools or --tools entry "doesn't look like a complete rule." It explains why MCP servers stop loading when an admin adds managed-mcp.json mid-session. It also fixes rules added in /permissions looking saved when the settings file couldn't be written, and fixes a settings file saved with an error being skipped silently. The latest tag points to 2.1.326, while stable still points to 2.1.311. CHANGELOG · npm
OpenCode Browse hit the Hacker News front page on November 16 with about 285 points by this morning. It is an MCP server plus a Chromium extension that lets coding agents drive your real browser, with your tabs, logins and extensions. Agents work in background tab groups. They hand the tab back to you for a login, a 2FA prompt or a payment, and an allowed-sites list keeps them on named domains. The product page carries no launch date, so we date it by the HN post. The thread split on whether permission prompts inside a harness help at all, and one commenter argued for an outside sandbox instead. OpenCode · HN
New tools and features worth actually trying
LangSmith's langsmith setup. The November 9 to 16 changelog says it installs the LangSmith skills and the LangSmith MCP server into Claude Code, Cursor and Codex on your machine "with no API key written to any agent config." Honest tradeoff: it only helps if you already trace with LangSmith, and the changelog does not say where the key does live, so check that before you run it on a shared machine.
A verify skill, Claude Code team style. Put a skill named verify in .claude/skills/ next to the code it covers, have it start the app and drive one real user flow, and ask Claude to stress-test the skill itself until it is boring and reliable. Honest tradeoff: an end-to-end smoke test is slow and flaky until you invest in it, and the post itself warns that flaky tests teach an agent to retry rather than fix.
OpenCode Browse's allowed-sites list. If you let an agent use your logged-in browser at all, set the allowed sites first and use the handoff for every login and payment. Honest tradeoff: the agent still sees everything your session can see on those sites, and an allowlist is not a sandbox.
LangSmith sandbox functions. You can now deploy a sandbox snapshot as an HTTP function with min_instances, max_instances and idle_timeout_seconds, and roll back by redeploying an older snapshot. Honest tradeoff: it is new, and it ties a piece of your serving path to LangSmith's platform and pricing.
Trending AI repos on GitHub today
Read from Trendshift at about 07:20 ET on November 17. Its numbers are live momentum scores, not star totals, and ranks drift through the day (a later read had several of these swapped by a place or two). Stars, licenses and releases below come from the GitHub API, the LICENSE file and each repo's releases feed.
- mvanhorn/last30days-skill (#12): an agent skill that searches Reddit, X, YouTube, HN, Polymarket, GitHub and more for the last 30 days and has an agent judge write one brief. Why now: v4.3.0 landed November 16. MIT, about 68.6k stars. Caveat: several sources need your own keys or logged-in browser sessions, and it ranks by engagement, which is not the same as accuracy.
- mdiri-ai/SoduWork (#14): a self-hosted, multi-harness workspace for teams and their agents, where a request ends only when the person who asked signs off. Why now: v1.13.17 on November 16. MIT, about 1.5k stars.
- ethanplusai/sellbird (#22): an outreach agent (formerly Harvey) that finds companies and decision makers, writes cold email and handles replies, running on your Claude or ChatGPT plan. Why now: renamed in v0.4.0 on November 6. AGPL-3.0, about 6.1k stars. Caveat: by default nothing sends until you approve it, and the README also points to a hosted paid version.
- dream-num/univer (#10): an embeddable spreadsheet, document and slide SDK pitched for people and AI agents. Why now: v1.5.2 on November 16. Apache-2.0, about 28.7k stars. Caveat: the default branch is
dev, notmain. - p-e-w/heretic (#11): training-free model editing that strips refusals and cuts sycophancy and slop. Why now: v2.1.0 on November 11. AGPL-3.0, about 41.2k stars, default branch
master. Caveat: removing refusals removes safety behavior along with the annoying kind. - EL4CTEO/blender-skills (#17): 28 Agent Skills covering the Blender pipeline, from modeling to export, for Claude Code, Codex, Gemini CLI, Copilot and Cursor. Why now: v1.3.1 on November 13. MIT, about 1.7k stars. Caveat: the README targets Blender 5.2 LTS or newer.
- huggingface/pybot (#20): a chatbot with no weights, only if-else branches, regex and lookup tables, written by a swarm of more than 1,000 agents. Why now: it is a strange, public test of what agent swarms can build. Apache-2.0, about 450 stars, no releases. Caveat: its headline result (from 0% to 26% on SmolTalkBench, on par with SmolLM-360M-Instruct, per the README) is self-reported and graded by an LLM judge against reference answers.
- ccch1mneyyy/dsh-Go (#25): a phone app for DeepSeek Harness that runs the harness on the phone or remote-controls your computer. Why now: v1.1.0 on November 15. AGPL-3.0, about 25.6k stars. Caveat: a third-party project, Android only for now, and iOS users must build and sign it themselves.
What actually matters from today's signal
Track who can write the rule file. Claude Code's review rules come from the base branch, provided you protect it. Copilot customizations can be limited to plugins and managed settings, with a separate allowlist for where plugins come from. VS Code's new harness gives admins permission and sandbox controls they have to switch on themselves. Claude Code 2.1.326 even warns when a --allowedTools entry isn't a complete rule. For builders, the high-signal areas are clear: base-branch-only review and gate configs, plugin and marketplace allowlists for skills and MCP servers, managed harness settings with sandboxing turned on deliberately, and verify skills that run the real app before a commit.
The counter-signal is Optimizely. Every rule above protects a file from the agent, and the Optimizely attacker never touched an agent at all. Someone got write access to customers' Project JavaScript, and nobody yet knows how. At least 178 organisations then served the result under a valid certificate. That is the same pattern aimed at your agents: a skill, a plugin or an MCP server from a trusted marketplace is only as safe as everyone who can publish to it. OpenCode Browse makes the stakes concrete, since an agent driving your logged-in browser inherits whatever a poisoned page does. Locking the rule file is necessary. Watching what the trusted file actually contains, every day, is the half most teams still skip.
Source access notes: Read Tuesday November 17 from about 07:13 to 07:50 ET. WebFetch was not available as a direct tool, so pages were read through TailFinder's web reader on the linked Mac. The cloud shell reached raw.githubusercontent.com and the npm registry but not vendor blogs, Trendshift or GitHub pages. The Hacker News Algolia API returned JSON that the reader refused as non-HTML, so the HN front page was read directly. Vendor pages checked: openai.com/news (latest items Nov 16, already covered in yesterday's afternoon brief), anthropic.com/news (nothing after Nov 10), blog.google, blog.cloudflare.com (nothing after Nov 11), Microsoft Foundry blog (Ignite runs Nov 17 to 20; nothing new yet at read time), mistral.ai/news (nothing after Nov 5), huggingface.co/blog, GitHub Changelog, LangSmith changelog. Not checked this run: x.ai, Microsoft Agent Framework blog, DeepMind, the Codex changelog and Product Hunt. Repo figures come from a Sonnet 5.5 verification subagent that read the GitHub API, LICENSE files and releases.atom through the Mac. Univer's README and LICENSE were read from main though its default branch is dev. No Optimizely statement was found, so all incident figures are attributed to Report URI. A Sonnet 5.5 adversarial pass caught: the thesis overstating GitHub's setting as "managed plugins only" (the changelog says plugins or managed settings, and plugin-only is not an allowlist); the Claude Code base-branch rule losing its protected-branch condition; 178 organisations stated as a total rather than a floor; the 2.1.326 warning scoped too broadly; "written" dropping "and reviewed"; and a misleading pybot benchmark caveat. All fixed. It also flagged that Trendshift ranks had drifted on a later read; ranks here are from our 07:20 ET read.