Independent AI intelligence Two editions daily · ET
Fervor AI

AI Trending Briefing · November 19, 2026 · morning edition

On November 18 agents moved from asking to staying connected, as GitHub put event streams into gh and added eight MCP connectors and Claude Code let MCP-backed policy hooks block on failure, so what happens when a check breaks is now the setting to audit.

Gemini 4 NeonGitHub Event StreamsGitHub Copilot connectorsClaude CodeRuby2ShellMicrosoft MAIagent-infrastructuremcpclaude-codeagent-securityfrontier-models

Trending AI Briefing: Thursday, November 19, 2026 (morning ET)

Most agent loops still work like a kid in a back seat asking "are we there yet?" every thirty seconds. November 18 chipped away at that. GitHub gave the gh CLI a way to hold a pull request subscription open and wait, Copilot picked up eight more services it can stay plugged into through verified MCP servers, and Google pitched Gemini 4 Neon as an orchestrator that directs parallel subagents. The more an agent stays connected, the more it matters what happens when the check in front of it breaks, which is why the quietest item of the day, a Claude Code release that lets MCP-backed policy hooks block when they fail, might be the one to act on first.

What's hottest in AI news right now

Google released Gemini 4 Neon on November 18, priced at $3 per million input tokens and $15 per million output, with a 1 million token context window. Google says it scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level (third overall by Google's count), 65.2% on Terminal-Bench 4.1 and 58.6 on the Artificial Analysis Coding Agent Index. The builder detail is the role Google assigns it: a "fast, high-judgment Orchestrator directing parallel Hydrogen subagents" inside Google Antigravity. Cached input drops to $0.15 per million. It is live in the Gemini API, Google AI Studio and Antigravity for paid Pro and Ultra subscribers, and GitHub added it to Copilot the same day. The catch: every number in the post is a leaderboard Google chose, and a separate "Neon for cyber" variant goes only to vetted defenders through the Fairwind Program "in the coming days." (Google, GitHub changelog)

Ruby shipped 4.0.10, 3.4.13 and 3.3.14 on November 18 to fix seven memory-safety bugs in its regex engine and string case conversion, and Veria Labs says one of them is a pre-authentication remote code execution on stock Rails. The Ruby advisory says the team knows of proofs of concept that reach code execution for some of the seven, and CVE numbers are still pending. Veria's write-up, branded Ruby2Shell (GHSA-34m4-m9fh-gqp5), traces it to a one-byte heap overflow: upcasing ß in most single-byte encodings writes two bytes while the loop checks room for one, and the bug has sat there since Ruby 2.4.0. Veria lists Rails and classic-style Sinatra as reachable in default configuration, and as of 12:00 UTC November 19 it marks GitLab 19.6.0, GitHub Enterprise Server 3.22.5 and Chatwoot 4.20.0 as bundling an affected Ruby. Veria plans to publish its proof of concept in a week. UTF-8 strings take a different code path, so the exposure depends on legacy encodings reaching a case conversion. (Ruby advisory, Veria Labs)

GitHub CLI v2.105.0 added gh event listen on November 18, the command-line face of the Event Streams preview. Point it at a pull request URL and the events you care about (a submitted review, a state or head change, a merge conflict, a finished check suite) and it opens the stream and prints each event as a line of JSON. By default it exits after the first event, which is exactly what an agent loop wants. --follow, --count and --timeout cover longer waits. GitHub also changed gh pr create, gh pr view, gh run watch and gh api to suggest gh event listen instead of polling, so agents will find it without being told. The catch: Event Streams itself is still behind a waitlist, and the commands "may change." (GitHub changelog)

Copilot gained eight more connectors on November 18: Box, Buildkite, Calendly, ClickUp, Mixpanel, Supabase, Webflow and Wix. Connectors are GitHub-verified MCP servers, joining more than 25 existing ones such as Slack, Figma and Sentry, and you connect once to use them across the Copilot app, CLI (/mcp, then the Connectors tab), VS Code and github.com. GitHub's own example is checking a Supabase schema before writing a migration. The catch sits in the availability note: on Business and Enterprise an admin must enable editor preview features, and an enterprise MCP allowlist that names connectors one by one has to be updated by hand. (GitHub changelog)

Claude Code 2.1.328 reached npm at 23:09 UTC on November 18 and added "onFailure": "block" for MCP tool hooks. Hooks are the user-defined checks Claude Code runs at points like "before a tool call." Command and HTTP hooks have had the option since v2.1.295, and the hooks docs spell out why it exists: on most events a hook that fails or times out lets the action go ahead, so "a policy hook with a wrong path or a crashing script then lets everything through." The release also adds auto mode for Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud's Agent Platform and Foundry, plus a long list of self-hosted runner flags. The catch: the hooks docs still list the field only for command and HTTP hooks, so the MCP version is changelog-only for now, and latest is 2.1.328 while the npm stable tag still points at 2.1.311. (changelog, npm, hooks docs)

Microsoft AI released a batch of models at Ignite on November 17, led by MAI Code 2 and MAI Code 2 Flash in GitHub Copilot. Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price, averaged over four benchmarks it picked, and that MAI Cyber 2 Flash is priced 86% lower on output tokens than MAI Cyber 1.1 Flash. Unusually, Microsoft says what the coding models are bad at: they are "explicitly not yet tuned" for fundamental science or very long-horizon agent work. All comparisons are Microsoft's own. (Microsoft AI)

New tools and features worth actually trying

gh event listen. If an agent babysits pull requests, replace its polling loop with gh event listen <pr-url> filtered to submitted reviews, then pipe the JSON through --jq. Honest tradeoff: you need Event Streams waitlist access first, and the command surface is labeled preview.

"onFailure": "block" on your Claude Code policy hooks. Add it to any PreToolUse hook you treat as a gate, then test by pointing the hook at a missing script and asking Claude to run ls; the docs say the call should be blocked with blocking because onFailure is "block". Honest tradeoff: a flaky hook now stops work instead of being ignored, the documented path is command and HTTP hooks (the MCP hook version is in the 2.1.328 changelog but not yet the docs), and the field does nothing on Stop, SubagentStop, TaskCompleted, TeammateIdle or background hooks.

aac run from Bitwarden Agent Access. It fetches one credential from your vault over an encrypted tunnel and passes it to a child process as environment variables, for example aac run --domain example.com --env DB_PASSWORD=password -- psql, so the secret never prints to the agent's transcript. Honest tradeoff: early preview, protocol subject to change, and the relay server in the repo is for demos only.

A ten-second Ruby check. Run ruby -v on every host and container image that serves Rails or Sinatra and upgrade anything below 4.0.10, 3.4.13 or 3.3.14. Honest tradeoff: self-hosted GitLab and GitHub Enterprise Server ship their own Ruby, so upgrading the system interpreter does not fix them; you wait for the vendor release.

Trending AI repos on GitHub today

Read from Trendshift's daily board at about 07:15 ET; its figures are momentum scores, not star totals, so star counts below come from shields.io (rounded) and licenses from each repo's LICENSE file.

  • BaaRai/ARTEX (#2): a self-hosted pentest platform where a planner agent dispatches worker agents that run real security tools and tie findings to evidence. Why now: v0.6.0 on November 12, and Cloudflare published countermeasures against orchestrators like it on November 13. AGPL-3.0 per the LICENSE file, about 11k stars. Caveat: it is an English-translation fork whose upstream is no longer public, and the README mentions extra "usage restrictions" that the unmodified AGPL file does not contain.
  • simplexai-labs/linger (#5): a personal agent that runs a local model on your Mac or PC. Why now: desktop-v0.1.4 shipped November 19. About 1k stars. Caveat: the license is a modified Apache-2.0 that requires written permission to offer it as a hosted service, the benchmarks are self-run, and the default Docker login is owner@linger.local / test1234.
  • aoci-spec/aoci-code (#7): a Git-versioned index of your codebase and database schema that your coding agent writes and an MCP server serves back at session start. Why now: v0.2.0-rc2 on November 17. FSL-1.1-MIT (source-available, converts to MIT after two years), about 2.5k stars. Caveat: index entries are model-written and the validator checks structure, not accuracy.
  • bitwarden/agent-access (#8): an open protocol, CLI (aac) and Rust SDK that hands an agent one credential from a password manager over a Noise-encrypted tunnel instead of the whole vault. Why now: climbing the board with no new release since v0.11.0 on March 21. Apache-2.0, about 580 stars. Caveat: the connect path prints the password as JSON, so it lands in whatever transcript runs it; the README itself steers LLM use to aac run.
  • felipebasurto/unslop-ui (#14): an agent skill that strips AI-style decoration (eyebrow labels, status dots, decorative icons) from UI copy and charts. Why now: v0.5.1 on November 17. MIT, about 400 stars. Caveat: its "31 tells to 0" chart counts regex hits on fixtures written to contain tells.
  • trypostit/trypost (#17): a social media scheduler built for agents, with an MCP server and REST API that publish to 13 networks. Why now: v2.0.1 on November 1. AGPL-3.0, about 1.4k stars. Caveat: the README gives no self-hosting runtime or environment variables while promoting a hosted cloud.
  • firecrawl/firecrawl (#20): search, scrape and crawl the web into LLM-ready Markdown or JSON. Why now: v2.11.797 on November 18. AGPL-3.0, a very large following. Caveat: the README's "96% of the web" and P95 3.4s figures carry no methodology on the page, and the cloud has features self-hosting lacks.
  • storytold/craft-anything (#23): an early-alpha Rust engine that studies an existing site, app or game, then has a coding agent build a new version. Why now: the board's biggest momentum score today. Dual Apache-2.0 and MIT, about 38k stars; release status not verified this run. Caveat: needs Rust 1.97+ and a signed-in coding agent, and the README says most agents other than Claude Code are "built, not yet run live."

What actually matters from today's signal

The trend to track is the agent that stays attached: to a pull request through an event stream, to Supabase or Buildkite through a verified connector, to a vault through a credential tunnel, to a fleet of subagents as their orchestrator. For builders the highest-signal areas this week are event-driven agent loops (drop polling where you can get the waitlist), connector scope reviews (a "verified" server tells you who built it, not what your token can touch), credential handoff that keeps secrets out of transcripts, and fail-closed policy hooks.

The counter-signal is that most of the guardrails stay opt-in. Claude Code still lets a crashed hook through unless you set a field. Copilot's connector allowlist only protects you if someone keeps it current. Bitwarden's own README warns against typing credentials into agents at all. And the day's most practical security news had nothing to do with AI: Ruby2Shell is a one-byte bug in code that dates back to Ruby 2.4.0 (December 2016), and GitLab and GitHub Enterprise Server were still shipping the affected Ruby on Thursday morning. Anthropic priced its IPO the same day, and that will dominate the timeline. The patch queue deserves the attention more.


Source access notes: Read Thursday November 19 from about 07:11 to 07:45 ET. WebFetch was not a direct tool in this session, so pages were read through TailFinder's web reader on the linked Mac. The Hacker News Algolia API was unreachable from both the cloud workspace and the Mac, so the HN front page stood in. OpenAI's GPT-6.2 and "Research acceleration" pages returned HTTP 500 and are not used here. The Claude Code changelog and npm registry were read directly from the Mac. Repo figures came from a Sonnet 5.5 verification subagent using cache-busted shields.io, raw README and LICENSE files, and releases.atom; craft-anything's release feed was rate-limited. A Sonnet 5.5 adversarial pass ran on the draft. Applied: softened "several" to "some" for the Ruby advisory's proof-of-concept wording, and flagged that the hooks docs still list onFailure only for command and HTTP hooks. Held: it reported Veria's page as last updated 04:35 UTC with an older product list; the page read in this session at about 07:20 ET showed "Updated 19 November 2026 at 12:00 UTC" with Mastodon, OpenProject and Discourse already fixed, so the newer read stands. Held: it could not see npm's time map; the packument read from the Mac gives 2.1.328 at 2026-11-18T23:09:39Z.