Skip to the page
Tecyfy

9 November 2026 to 15 November 2026

Gemini 4 Hydrogen ships; Claude Code's sandbox asks first

AI and data news from 9 November 2026 to 15 November 2026: Claude Code's sandboxed Bash commands ask for permission again; GitHub Copilot drops 6 models on 14 December; Codex exec keeps MCP servers off in untrusted folders, and 16 more.

FeatureAnthropic

Claude Code's sandboxed Bash commands ask for permission again

Since Claude Code 2.1.325 on 13 November, Bash commands that run in the sandbox follow the normal permission flow by default instead of running without a prompt. If you relied on the old behaviour, set sandbox.autoAllowBashIfSandboxed: true; an organization that turns the sandbox on through managed settings sets it in its highest-ranking managed source. The same release makes read-only git commands ask first in a repository you have not trusted, makes scheduled tasks and /loop ask before running a skill that declares allowed-tools or hooks (add Skill(name) to permissions.allow to run it unattended), and closes gaps where some compound commands, or a Bash allow rule, could run a program without asking.

Where it came from

ReleaseGitHub

GitHub Copilot drops 6 models on 14 December

GitHub will deprecate 6 models across every Copilot experience on 14 December: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3. Their replacements are Claude Fable 5.2, Claude Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon and Grok 4.7; Kimi K3's is still to be announced. Individual subscribers on annual plans keep Claude Haiku 4.5. If your workflows or integrations name one of these models, move them before that date; Business and Enterprise admins who turned off default model enablement must turn the replacements on in model policies.

The idea behind it

FeatureOpenAI

Codex exec keeps MCP servers off in untrusted folders

Codex 0.169.0, released on 13 November, makes codex exec keep a project's configuration and MCP servers disabled in a directory you have not trusted. A repository you clone and run codex exec in can no longer start its own MCP servers that way until you trust it. Update the Codex CLI if you script it over code from other people.

Where it came from

FeatureDatabricks

Databricks Genie ZeroOps watches jobs and proposes fixes, in beta

Genie ZeroOps, in beta since 12 November, monitors jobs, pipelines and tables in the background. It finds failures and data quality problems, finds their root cause and proposes a fix.

The idea behind it

FeatureAnthropic

Claude Code now assumes a 1M window for models it doesn't know

Claude Code 2.1.323, on 12 November, treats a model ID it does not recognize as having a 1M token context window, up from 200K. If you run it on another vendor's model with a smaller window, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before the session fills, or it can fail on every turn until /clear. The release also adds Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools, and removes the v1 MCP client: MCP_SDK_GENERATION=v1 now runs the v2 client in a compatibility mode.

Where it came from

FeatureAnthropic

Claude Managed Agents can run entirely on your own servers

Since 12 November, Anthropic offers self-hosted Claude Managed Agents: you run the whole service, not only the sandboxes, on your own infrastructure, and agents, sessions and session history are stored there. Model requests go to the provider you choose, Anthropic by default. Access is on request through your Anthropic account team.

Where it came from

FeatureAnthropic

Claude can join a Google Meet call, in beta on Team and Enterprise

Claude Meetings began rolling out in beta on 12 November for Team and Enterprise plans, in Claude and Claude Desktop. You add Claude to a Google Meet call as a visible participant that takes notes, answers questions and builds docs, whiteboards or slides during the call.

Where it came from

ReleaseGoogle

Gemini 4 Hydrogen is generally available, from $1.50 per million

Google made Gemini 4 Hydrogen (gemini-4-hydrogen) generally available on 12 November, on the free tier and at $1.50 per 1M input tokens and $7.50 per 1M output tokens on the paid tier. It adds a fourth thinking level, xhigh, beside low, medium and the default high, and Google publishes a migration page for moving from Gemini 3.8 Flash. The same week it reached Databricks Unity Gateway through Foundation Model APIs and, rolling out gradually, GitHub Copilot Pro, Pro+, Max, Business and Enterprise, billed at the provider's list price.

FeatureDatabricks

Databricks can deny access to MCP servers, agents and skills by policy

Since 12 November, in beta, attribute-based DENY policies in Unity Catalog can apply to model services, model provider services, MCP services, agent services and skills, and can be attached directly to a single table, volume, function, model or secret. A governance team can now block a group from an MCP server or an agent with the same rules it already uses for tables.

ReleaseGoogle

Gemini Decision answers yes-or-no and category questions

Google released Gemini Decision (gemini-decision-1) and a Decisions API on 11 November, for binary, categorical and rubric decisions across text, images, audio, video and PDFs. Its decisions come back with zero output tokens. Prices are on Google's pricing page.

Where it came from

FeatureDatabricks

Claude and ChatGPT can query Databricks through Genie One

Since 11 November a Databricks Genie connector for Claude and a Genie plugin for ChatGPT connect each assistant to the Genie One MCP. An admin sets it up once for the team, and each person signs in with their own Databricks account to ask questions of their data. Databricks' context-based ingress policies can now let ChatGPT and Claude traffic in as partner platforms.

FeatureDatabricks

Codex on Databricks can bill your own ChatGPT plan

From 11 November, Codex and the ChatGPT desktop app can use your own ChatGPT Go, Plus, Pro or Enterprise subscription through Databricks Unity Gateway. Usage is billed to that subscription, while Unity Gateway still governs access.

ReleaseGoogle

Gemini 4 Carbon is generally available at an introductory price

Google made Gemini 4 Carbon (gemini-4-carbon) generally available on 10 November, at $2 per million input tokens and $10 per million output until 30 November 2026. Google's changelog does not say what it costs after that date, so check the pricing page before you budget past November. The same day it reached GitHub Copilot and Databricks Unity Gateway.

ReleaseAnthropic

Claude Fable 5.2 is out, and costs 20% less than Fable 5.1

Anthropic released Claude Fable 5.2 on 9 November at $8 per million input tokens and $40 per million output, down from $10 and $50 for Fable 5.1. It has a 1M-token context window, 128k output tokens and adaptive thinking that is always on, and runs on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, as well as GitHub Copilot and Databricks Unity Gateway. Claude Code 2.1.321 makes it the default Fable model and moves saved Fable 5.1 picks over at once on Pro and Max, and on Team and Enterprise once the organization allows it; /model claude-fable-5-1 switches back. Fast mode on the Claude API costs $16 and $80 for up to 2.5x the output speed, and like Fable 5.1 the model needs 30-day data retention.

FeatureAnthropic

Moving to Fable 5.2 can break 3 kinds of API request

Code that edits the system prompt, tools or an earlier message and then sends back a Fable 5.2 thinking block now gets a 400 error, on every account. A request that declares the older computer_20251124 tool also fails everywhere but Amazon Bedrock; switch to computer_toolset_20260801. With the advisor tool in beta, a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor. And if a conversation moves from Fable 5.2 back to Opus 5.5 or Fable 5.1, the API drops its thinking blocks, so check your code before you change the model name.

Where it came from

FeatureAnthropic

Claude Code 2.1.321 stops MCP refresh tokens leaking on a redirect

Claude Code 2.1.321, out on 9 November, fixes MCP OAuth requests that sent secrets such as refresh tokens to another host when a server redirected them. The same release fixes a read-only Bash command that ran outside the sandbox without a prompt, and deny rules that missed the file a cp or mv into a folder writes. /bin/cp and /bin/mv now get the same path checks as cp and mv. Update Claude Code if you connect it to MCP servers that sign in with OAuth.

Where it came from

IncidentGPUStack

Attackers are using a GPUStack flaw to take admin control

CISA added CVE-2026-102895 in GPUStack, an open-source GPU cluster manager, to its list of exploited flaws on 9 November. Before version 2.2.0, GPUStack gives admin rights to any request that looks like it comes from the same machine, which a reverse proxy on that host or a forged X-GPUStack-Real-IP: 127.0.0.1 header makes true; attackers can then create admin API keys and run code on worker hosts. The advisory says it was exploited against a 2.1.1 server. Upgrade to 2.2.0; until then set force_auth_localhost=true, strip that header at your proxy and keep the API off untrusted networks.

IncidentFireworks AI

Fireworks AI says customer model weights and data were reached

Fireworks AI updated its incident bulletin on 9 November: on 24 October an unauthorized party used a breached third-party cloud provider to reach Fireworks development machines. It says resource metadata, billing details and, for certain customers, model weights and datasets were affected, and secrets for a limited set of accounts; it saw no impact on Inference, the Training SDK or Nexus, and card numbers are not stored. Fireworks is emailing the customers whose data was affected with the steps to take; if you use Fireworks, check for that email, and contact security-notice@fireworks.ai with questions.