Radar, November 14, 2026: ten signals from X on llama.cpp 1.0, the Cursor incident and the DeepSeek API

By Ivo Donker · Saturday November 14, 2026 — compiled by Claude (AI, Anthropic) from X search briefings by Grok Bot (AI, xAI) and the llmnet team's X trend monitor

All daily signals · Markdown version

In short: The standout signal of November 14: llama.cpp and ggml reached version 1.0.0, so breaking API changes now get a new major version. Alongside it, ten short signals on how a leaked vendor key at Cursor led to a fake extension, DeepSeek's app functions in the API, a security fix in Claude Code and agents in Antigravity and Perplexity, each with its source on X.

Illustration: a radar screen on a desk, with a llama with gears on a pedestal and a panel of switches on the left, and a leaking key over a bucket and a window with tabs on the right.

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

1.llama.cpp and ggml are v1.0.0: breaking API changes now get a major

Confidence: high (release and HF blog read)

The first stable major version (November 13) promises that a break in the public C API or ABI gets a new major. 35 deprecated functions are gone; new are an FP8 KV cache, a MoE expert cache with a CPU share and a safer llama-server. On November 14 Ollama was still on v0.40.16, without the 1.0 API.

Source on X: @ggerganov · @nikwen_de · GitHub, release v1.0.0

Context in the briefing: Open source in the November 14 briefing

2.Cursor: fake Claude Code extension started with a leaked Fireworks key

Confidence: high (blog read)

A key Cursor issued to Fireworks was exposed there. An attacker used it to copy 2025 agent traces from sessions without Privacy Mode, found credentials in them, and on October 29 placed fake versions 2.1.312 and 2.1.313 of the Claude Code extension in the marketplace cache, for about 11 hours, reaching at most 5,000 users.

Source on X: @cursor_ai · Cursor-blog

Context in the briefing: Agent security in the November 14 briefing

3.DeepSeek puts search, files, link reading and speech in the API

Confidence: high (docs and pricing page read)

Since November 13 deepseek-flash (V4.1-Flash) runs server-side tools: web search with citations, search during reasoning, files, URL reading and a safety check, across Messages, Responses and Chat Completions. A search and a page fetch each cost $0.75 per thousand. Public apps must enable the safety check.

Source on X: @deepseek_ai · DeepSeek API-docs

Context in the briefing: DeepSeek in the November 14 briefing

4.Claude Code v2.1.325: security fix in zstd decoding, and themes via mods on desktop

Confidence: high for the release (read); mods from X only

The November 13 release opens with a fix for a security issue in decoding zstd-compressed responses; it names no CVE or affected versions. Also: subagents with their own model stay on that model, and sessions through third-party gateways compact again. @ClaudeDevs also announced mods for custom themes in the desktop app.

Source on X: @ClaudeDevs · GitHub, release v2.1.325

Context in the briefing: Claude Code in the November 14 briefing

5.Gemini 4 Hydrogen in Antigravity, with an Orchestrator mode

Confidence: medium (official posts on X)

@googledevs says Gemini 4 Hydrogen is in Google Antigravity, with an Orchestrator mode in which a heavier model plans and a lighter model executes. With /run-workflow the agent writes a Python script that runs subagents in parallel or in sequence. We did not try the mode.

Source on X: @googledevs · @sundarpichai

Context in the briefing: Google in the November 14 briefing

Read also: Gemini 4 Hydrogen · Gemini 4 Carbon

6.Perplexity Teammates in the Mac app for Max subscribers

Confidence: medium (official post on X)

@perplexity_ai says Teammates are available in the Mac app for Max subscribers: agents that take on work alone or together, such as QA, customer service, research and marketing. Before this we only knew Teammates from a TestingCatalog post. We did not read a product page or changelog.

Source on X: @perplexity_ai

7.Hermes Agent: inference via Modal and a Home Assistant plugin

Confidence: low to medium (X only; Home Assistant claim from a fan account)

Teknium of Nous Research says Modal is available as an inference provider through a verified plugin in Hermes Agent. A fan account writes that Home Assistant became an official plugin, letting the agent control lights, heating and blinds. If you let an agent run your home, limit which devices it may control.

Source on X: @Teknium · @HermesWatcher

8.Cloudflare open-sources Kitesurf, a lightweight browser engine for agents

Confidence: medium (official changelog account on X)

According to @CFchangelog, Kitesurf is open source under Apache 2.0: a lightweight browser engine for AI agents that runs on Workers. A user points out that a self-hosted instance runs without authentication by default. We did not read the repository itself.

Source on X: @CFchangelog

Context in the briefing: Open source in the November 14 briefing

9.ChatGPT Wallet: agents pay through Dots, first for Pro in the US

Confidence: medium (TestingCatalog and TechCrunch on X)

According to @testingcatalog, ChatGPT Wallet is rolling out on Dots for Pro subscribers in the US: agents can make purchases without real card details, with a security check per purchase. We did not read an OpenAI announcement; reports of an EU rollout were not confirmed.

Source on X: @testingcatalog · @TechCrunch

Context in the briefing: ChatGPT Wallet in the November 14 briefing

10.Unconfirmed: layoffs at Grok Imagine and a Grok 4.8 Beta

Confidence: low (X and secondary sources only)

@aatilley reports that the team behind Grok Imagine (about 60 people) was laid off; other outlets all point to The Information, and xAI confirmed nothing. Another account writes that Grok 4.8 Beta is coming as the first 2.5-trillion-parameter model. Both are rumors.

Source on X: @aatilley · @testerlabor

Context in the briefing: Unconfirmed in the November 14 briefing