AI news, November 16, 2026: MCP flaw in mcp-server-git, Qwen4-Max, open models, Anthropic IPO
Short answer: The main stories of November 15 to 16: the MCP reference server mcp-server-git had a serious vulnerability fixed in version 2026.11.11, Alibaba offers Qwen4-Max through QwenCloud, Meituan and Thinking Machines released open weights and Anthropic is starting an IPO for November 19. Claims about a leaked Claude Code prompt, Hermes with a Gemini subscription and Grok 4.8 are unconfirmed.

Illustration by Claude (AI), drawn with code: a Git branch with a cracked shield (MCP flaw), a stack of model blocks with a chip (Qwen4-Max and open models) and a price line with a bell (IPO).
The project behind the MCP reference servers published a security advisory about mcp-server-git on November 11. The server ran Git in a client-chosen directory without neutralizing that directory's own configuration. As a result, even a read-only call such as git_status could start a program named in the .git/config file (through core.fsmonitor, or diff.external in the diff functions), with the server's privileges. The advisory lists versions before 2026.11.11 as affected, rates the severity High (CVSS 7.8) and names 2026.11.11 as the fixed version. News roundups report five more flaws in the reference servers in the same release (including the fetch and filesystem servers) and two vulnerabilities in VS Code before version 1.145.3. Those were read only through the roundups.Confidence: high for mcp-server-git: the GitHub advisory was read directly; medium for the other flaws, VS Code and the uvx analysis: only through news roundups
What this means: if you run mcp-server-git, for example for a coding agent, update it, and do not open repositories of unknown origin with an agent that has Git tools. Pinning an MCP server version only helps if its dependencies are pinned too; a Corelight analysis shows that bare uvx calls resolve dependencies again at every rebuild. Background is in our overview of AI and cybersecurity and in our overview of MCP servers and agent tooling.
Sources: GitHub-beveiligingsmelding GHSA-6hqx-gqrx-c5r2 (mcp-server-git) · VS Code-melding GHSA-4x35-6qr8-ww48 (niet gelezen) · Weekoverzicht Oakgen (VS Code, Corelight-analyse)
Alibaba Cloud shows Qwen4-Max on QwenCloud as the new flagship of the Qwen family, with text, image and video input and text output. The model page lists a context window of 1 million tokens, more than 390,000 tokens of output and a price of 1.50 dollars input and 4.50 dollars output per million tokens, with separate cache rates. According to the page the API works with the OpenAI and the Anthropic format. News reports say the model launched on November 16 after a preview last week and that smaller open-weight variants follow later; the model page itself gives no launch date.Confidence: medium: prices and specifications read on the model page; launch date and performance figures only through news reports and as stated by the company
What this means: the prices are below those of many Western frontier models, but figures on speed and cost per task come from the company itself. Test the model on your own tasks before putting it in a workflow, and read the data terms if you send company data. More about the preview is at Qwen4-Max.
Sources: QwenCloud-modelpagina Qwen4-Max · Sulat over de lancering (16 november)
According to The Decoder, Meituan released the weights of LongCat-2.5 on November 15: a mixture-of-experts model with 1.6 trillion parameters of which about 48 billion are active, 1 million tokens of context and an MIT license. Meituan says it was trained entirely on Chinese chips and itself reports a score of 88.4 on Terminal-Bench 2.1. On November 12 Thinking Machines put the weights of Inkling 2 Small on Hugging Face under Apache 2.0 (594 billion parameters in total, 17 billion active, at least about 500 GB of GPU memory). Perplexity released Decider v1.3, a routing and classification model under Apache 2.0. All benchmark figures come from the makers.Confidence: medium to low: read only through news roundups; the model cards themselves were not read back and all benchmark figures are the makers' own
What this means: these models are too large for a home server and matter mostly to those who want to host themselves or assess the license and the training chain. For small setups Perplexity Decider v1.3 is more relevant. Read the model card and the license before deploying anything. Background on Inkling 2 Small is at Inkling 2 Small.
Sources: The Decoder over LongCat-2.5 (niet zelf gelezen) · ToolSift-overzicht 13 november (Inkling 2 Small) · ToolSift-overzicht 15 november (Perplexity Decider)
Several news outlets, including Decrypt, report that Anthropic has set the price range for its IPO at 70 to 78 dollars per share. 825 million shares are offered (801.4 million new and 23.6 million from existing shareholders). Pricing is scheduled for November 18 and trading on Nasdaq under the symbol ANT for November 19. A news roundup also reports that Opus 5.5 has been the default model in Claude on Pro, Max, Ultra and Team since November 12, with advice to use Medium effort; that was read only through the roundup.Confidence: medium: price range, share count and dates read in one article and reported by several outlets; the dates are planning; Opus 5.5 as default only through a roundup
What this means: the dates can still shift, so wait for the official pricing. This is news, not investment advice. For users nothing changes in the products; background on the filings is at the IPO and on the models at Claude Fable 5.2.
Sources: Decrypt: Anthropic zet IPO-prijsrange op 70 tot 78 dollar · Claude-artikel over Opus 5.5 als daily driver (niet zelf gelezen)
@CodeByNZ reports that Microsoft has published the open-source plugin TokenEconomizer, which learns from earlier agent sessions in a repository and loads fewer unused tools. It cites 18.4 to 30.9 percent lower cost on SWE-Bench for Claude Code, Codex, Copilot CLI and Pi; that figure comes from the post and not from our own measurement. @testingcatalog reports that in the Project settings of Claude Code you can now choose different models for the Coordinator and for Threads. According to the release notes, Codex CLI 0.169 leaves out project configuration and MCP servers in untrusted folders unless you pass --trust; @thsottiaux writes that the team may have fixed the 'model at capacity' error.Confidence: low to medium: posts on X read; the TokenEconomizer repository and the Claude Code documentation were not read back; the Codex release was read only through a roundup
What this means: the saving is in routing and in less context, not only in model choice. Read the TokenEconomizer repository before putting it on a working repository and measure cost over several sessions. If you use codex exec in scripts, check whether you now need --trust. See also Codex CLI 0.169 and our overview of agentic AI.
Sources: @CodeByNZ op X (TokenEconomizer) · @testingcatalog op X (Coordinator en Threads in Claude Code Projects) · @thsottiaux op X (model at capacity) · Codex rust-v0.169.0 op GitHub
@Tynric_carter claims that the full system prompt of Claude Code is in a public repository; Anthropic has not confirmed it. Chinese and Japanese posts report that Nous Research is talking to Google about official access for Gemini subscriptions in Hermes Agent; the post by @KOF97AI itself says the documentation only mentions API keys and the Nous Portal and that there is no release date. @kimmonismus writes that GPT-6.2 may not come and that OpenAI may be working on GPT-6.5; that is speculation. For Grok 4.8 a news roundup says no model ID, price or model card had been published on November 16. Two posts offer ready-made prompts for Grok Bot, OpenClaw or Hermes; those were not checked.Confidence: low: claims and speculation by accounts with reach without a primary source; unconfirmed
What this means: treat these as signals. Wait for an announcement from the maker or a primary source. Do not paste prompts or skills from unknown accounts into an agent with access to files or keys; that is a known prompt-injection pattern.
Sources: @Tynric_carter op X (systeemprompt Claude Code) · @KOF97AI op X (Hermes en Gemini-abonnement) · @kimmonismus op X (GPT-6.2) · @Av1dlive op X (ongecontroleerde promptset voor Grok Bot)
Primary sources: GitHub-beveiligingsmelding mcp-server-git · QwenCloud-modelpagina Qwen4-Max · Decrypt over de Anthropic-beursgang


