Skip to content
NLEN

Radar, November 16, 2026: ten signals from X on mcp-server-git, Teammates and Kitesurf

All daily signals · Markdown version

In short: The key signal of November 16: update mcp-server-git to 2026.11.11, because a malicious .git/config could start a program during a read-only call. Alongside it, ten short signals on agents with a standing role at Perplexity, Cloudflare's browser engine for agents, Qwen4-Max as a cloud API and a shorter clock at Project Zero, each with its source on X.

Illustration: a radar screen on a desk with four signals: a folder with a patch, three robots at desks, a browser window above cubes and a stack of coins with a meter.

Illustration: AI-generated (OpenAI image model via Codex): a radar screen with ten signals.

1.mcp-server-git: a .git/config could start a program during git_status

Confidence: high (advisory read directly)

The MCP reference server for Git did not neutralize the repository's configuration. Through core.fsmonitor or diff.external a read-only call could start a program with the server's privileges. Affected: everything before 2026.11.11, severity High (7.8). It only works when the .git directory itself comes along, as with an archive or shared folder; a normal clone does not. Updating is the only fix.

Source on X: @filipemotta · GitHub, GHSA-6hqx-gqrx-c5r2

Context in the briefing: The MCP flaw in the November 16 briefing

Read also: The mcp-server-git flaw · MCP and agent tooling

2.Perplexity Teammates: agents with a standing role, Max only for now

Confidence: high for the feature (help article read); growth figure from X only

A teammate gets a responsibility, plans its own tasks, has extra agents work in parallel and hands the result back for review. Research preview for individual Max subscribers, on the website; Claude Opus 5.5 by default. It uses your plugins with your permissions. Aravind Srinivas reports eleven times as many agent hours per week as at launch; that figure comes from X only.

Source on X: @AravSrinivas · @perplexity_ai · Perplexity, helpartikel

Read also: Perplexity Teammates

3.Cloudflare open-sources Kitesurf, and the engine has no authentication of its own

Confidence: high (blog and README read)

The browser engine for agents on Workers is on GitHub under Apache 2.0. It speaks CDP, so Puppeteer and Playwright can connect. Running it yourself needs a paid Workers plan. The README warns: without Cloudflare Access in front, anyone with the URL can browse through your account. According to Cloudflare it uses 4.7 times less memory than Chromium, but is slower per page.

Source on X: @CFchangelog · GitHub, cloudflare/kitesurf

Read also: Kitesurf open source · MCP and agent tooling

4.Qwen4-Max is cloud GA with an Anthropic-compatible endpoint, but without open weights

Confidence: high for price and specs (model page read); scores are the maker's

Alibaba put the model id qwen4-max live on November 16: 1 million tokens of context, image and video in, text out, and both the OpenAI and Anthropic Messages APIs. Internationally 1.50 dollars input and 4.50 dollars output per million tokens. Thinking is on by default. The smaller open Qwen4 models are announced and not yet on Hugging Face.

Source on X: @javierbits · QwenCloud, modelpagina

Context in the briefing: Qwen4-Max in the November 16 briefing

Read also: Qwen4-Max preview

5.Project Zero now discloses on day 90, whether or not there is a fix

Confidence: high for 90 days (policy page read in the 09:00 radar); later terms are a plan

Google Project Zero drops 90 plus 30 days. Under the new 'Full X' policy a report becomes public X days after it is filed, patched or not; X is now 90. For 2027 and 2028, 60 and 30 days are planned. The reason: language models speed up finding and exploiting bugs. Maintainers of agent tools get a shorter clock.

Source on X: @ShadowNewsHQ · Project Zero, beleid

Read also: AI and cybersecurity

6.Tencent Hy4 Flash: cheap MoE with 12 billion active parameters, no weights

Confidence: medium (only secondary sources citing the lab page)

Three Chinese news sites cite the lab page: 224 billion parameters, 12 billion active, 1 million tokens of context, API at 0.5 yuan input and 2 yuan output per million tokens. Free in WorkBuddy until November 30. Open weights are an intention; nothing is on Hugging Face. We could not read the lab page itself.

Source on X: @ShadowAIbbl1888 · Tencent Hy, labpagina

7.Codex: 'model at capacity' may be fixed, and another alpha without release notes

Confidence: medium (team post; tag read in the 09:00 radar)

@thsottiaux of OpenAI writes that the Codex team may have fixed the 'model at capacity' error. The new prerelease rust-v0.171.0-alpha.7 does not mention that fix; its body is one sentence and the commits are about plugin manifests and the TUI. Stable stays 0.169 with --trust for untrusted folders. If runs stalled on capacity, try them again.

Source on X: @thsottiaux · GitHub, rust-v0.171.0-alpha.7

Read also: Codex 0.169 and --trust · Codex news

8.Microsoft TokenEconomizer loads fewer unused tools into coding agents

Confidence: low to medium (X and news roundups; repository not read)

According to a post on X, the open-source plugin learns from earlier agent sessions in a repository which tools are needed and skips the rest. It cites 18.4 to 30.9 percent lower cost on SWE-Bench for Claude Code, Codex, Copilot CLI and Pi. That is the maker's figure. Read the repository before putting it on a working repository.

Source on X: @CodeByNZ

Context in the briefing: Token costs in the November 16 briefing

Read also: Skills and plugins

9.Claude Code Projects: a separate model for the Coordinator and for Threads

Confidence: low (X only)

TestingCatalog reports that Claude Code's Project settings now let you choose different models for the Coordinator and for Threads. That lets a large model plan while a cheaper one executes. We did not find Anthropic documentation on this.

Source on X: @testingcatalog

Context in the briefing: Token costs in the November 16 briefing

Read also: Claude Code news

10.Unconfirmed: Claude Code's system prompt and a GPT-6.2 that may not come

Confidence: low (X only)

One account claims the full Claude Code system prompt sits in a public repository; Anthropic has not confirmed it. Another writes that OpenAI is skipping GPT-6.2 and working on GPT-6.5. Both are claims without a primary source. Prompts from such a repository are not instructions to copy blindly.

Source on X: @Tynric_carter · @kimmonismus

Context in the briefing: Unconfirmed in the November 16 briefing