Cursor removes a malicious Claude Code extension, GitHub makes Copilot Fusion generally available, and more - AI Programming Weekly #75

In this edition, covering 25 October to 2 November:

• Cursor removed version 2.1.313 of Anthropic's Claude Code extension from its VS Code Marketplace after finding code that tried to install a persistent SSH backdoor; up to 5,000 users may have been affected.
• GitHub made Copilot Fusion, which selects and coordinates models for each task, generally available across all Copilot plans, and plans to add four million CPU cores after AI driven traffic contributed to outages.
• Google DeepMind announced Gemini 4 Carbon, which ranks first on the Artificial Analysis Intelligence Index v5 and goes first to cyber defenders and trusted partners.

This spot is open for 2027. One company puts its name on every AI Programming Weekly, which has 3,800+ subscribers. See what's included

Coding models and model releases

Google DeepMind announced Gemini 4 Carbon, which ranks first on the Artificial Analysis Intelligence Index v5 and leads cybersecurity benchmarks. It goes first to cyber defenders and trusted partners, priced at $2/$10 per million tokens at launch. Mistral released open weights for an improved one trillion parameter, natively multimodal version of Mistral Large 4 after additional reinforcement learning and planned safety testing. Mistral reports stronger coding and defensive cybersecurity performance and offers the model through its API, Mistral Studio, OpenRouter and Hugging Face under its Open Weight license. Gemini 4 Carbon · Mistral Large 4

Meituan released LongCat-2.5 on 2 November, a native multimodal model with 1.6 trillion total parameters that activates about 48 billion per token and is trained for agentic coding, visual coding and 3D tasks. It is available through LongCat's website and API, and Meituan's benchmark results have not been independently verified. Liquid AI released LFM3, six open weight 2B class models for text, vision, audio, omni and decision tasks on one backbone; the text model fits in about 1 GB at 4 bit. SemiAnalysis introduced GDPx, a benchmark across 11 domains including software engineering, chip design and cybersecurity, and reports that the leading model scored 12% overall. LongCat-2.5 · LFM3 · GDPx

Revolut moved about 2,000 engineers from Claude Code and Codex to a self hosted GLM model from Z.ai and says costs fell to about one tenth and coding time by 60%. Revolut

GitHub

GitHub made Copilot Fusion generally available across all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI and JetBrains IDEs. It selects and coordinates models for each task through single model, cascade or critique workflows, with administrator controls and three effort levels, and paid users get a temporary 30% discount on model costs. GitHub also plans to add four million CPU cores and 160 petabytes of high speed storage after AI driven code commits contributed to outages, move to a multiregion Azure deployment and give Copilot access to all GitHub Enterprise users without extra seat fees, while still charging for token use. Copilot Fusion · Four million cores

Entire, the startup of former GitHub CEO Thomas Dohmke, opened its Git compatible network for AI coding agents to all developers, free during preview. It replaces pull requests with Trails that store prompts and session logs; Microsoft's M12 invested in its $60 million first round. Entire

Agents in the developer workflow

At the Android Dev Summit, Google introduced Android Studio Rabbit 2 with support for third party coding agents including Claude, Codex, Junie and Antigravity through the Agent Context Protocol, plus agent tools for testing, code review, profiling and model cost comparisons. Factory introduced Droid 1.0, a coding agent harness for frontier models that adds programmatic tool calling through sandboxed JavaScript scripts, which can run hundreds of tool calls in one model request. Slack made the Slack Code API generally available, so AI agents can create and manage code channels, post progress updates and show code artifacts; GitHub integration will support review requests and merging pull requests from Slack. Android Studio · Droid 1.0 · Slack Code API

DeepSeek previewed a proposed super agent in which a central agent distributes tasks among multiple AI agents; in the company's tests it completed one task in under 16 minutes, against more than 29 minutes for a conventional approach, and DeepSeek flagged security and privacy risks. Epic Games says it uses Claude and Codex to develop code and features for Unreal Engine, including work on UEFN and UE6. Bloomberg introduced BQuant Assistant, which can generate and run approved Python and BQL code in BQuant Desktop and BQuant Enterprise. DeepSeek · Epic Games · BQuant Assistant

Goldman Sachs is limiting traders' and bankers' access to advanced AI coding tools, including Anthropic's Claude and GitHub, and keeping them mainly for engineers; quant strategists are exempt. Goldman Sachs

Upwork launched Upwork Now in public beta, which lets AI agents and clients reach freelancers who are ready to start work, and says AI skilled talent generated $112 million in third quarter gross service volume. Its MCP server has 20 times more weekly users than at its August launch. Upwork Now

StackRails launched a platform that maps code, infrastructure, APIs and engineering policies and gives AI coding agents that context before they write code. StackRails

Honeycomb opened early access to Personal and Team AI SRE Agents that use telemetry and MCP connectors to monitor deployments and investigate incidents. PagerDuty introduced Paige, an AI agent for incident investigation, and renamed its Operations Cloud as the PD Reliability Platform. Honeycomb · PagerDuty

Code and supply chain security

Cursor removed version 2.1.313 of Anthropic's Claude Code extension from its VS Code Marketplace after finding code that collected host details and attempted to install a persistent SSH backdoor. Up to 5,000 users may have been affected; Cursor shut down and rebuilt the marketplace, took the attacker's server offline and says no other extensions were tampered with. Northflank told customers to rotate all secrets after a breach flagged by Wiz on 30 October, and later said users reported unauthorised use of AI API keys stored on the platform. Wiz says at least nine AI developer platforms disclosed security incidents in October. Cursor · Northflank

Pluto Security reports that an operator used 28 Hugging Face repositories and a ModelScope mirror to distribute cryptocurrency miners and backdoors disguised as Qwen2 models, with more than 21,000 downloads by 25 October; the payloads could run when users loaded models with trust_remote_code=True. After the takedown, the campaign resurfaced in 31 repositories across seven accounts. Symantec says an unauthorized relay service installs genuine OpenAI Codex and Anthropic Claude Code tools, mainly at organizations in mainland China, then disables security sandboxes and replaces user credentials with its own tokens; no malicious activity has been observed. Hugging Face miners · Campaign returns · Relay service

Anthropic says an unreleased internal model found and chained flaws in Chrome, Safari, Android, Windows and Linux into proof of concept exploits, which were reported to vendors and patched; it did not disclose how much human help the model received. Veracode's October report says the average security pass rate for AI generated code rose from 56% to 62%, with Claude Opus 5.5 and GPT-6-Sol tied at 73% across 18 models. Cisco open sourced Project Quarry, a rule language and detection engine for multi step attacks across AI agent sessions, including tool calls and MCP interactions. Exploit chains · Veracode · Project Quarry

CrowdStrike and Lovable connected Lovable workspaces to Falcon Shield, so security teams can find publicly accessible apps, sensitive data and access issues in AI built projects. CrowdStrike and Lovable

Anthropic

Anthropic's IPO filing says the retail value of Claude used internally exceeded monthly employee compensation costs, excluding stock awards, in April. By September internal AI use had grown more than 250 times from January, and Claude wrote over 90% of the code added to Anthropic's software. The figures use retail pricing. Anthropic confirmed it updated Claude Code on Opus 5.5 with rules that steer it away from recreating specific existing games, while allowing original games. Engineering manager Felix Rieseberg says he no longer uses traditional coding interviews for his team and assesses candidates through résumés, portfolios and working prototypes. IPO filing · Game rules · Interviews

Anthropic made message threads available in public beta on the Claude Platform and through AWS, Microsoft Foundry and Google Cloud. Agents send only new messages after the first request, and Anthropic says this cuts median time to first token by about 35% for requests of 500,000 to 1 million tokens, with token usage unchanged. Message threads

OpenAI

OpenAI will combine ChatGPT and Codex usage allowances for Plus, Pro and Business plans starting in December, measured in tokens; Instant chats will not count toward the shared allowance. Cognizant formed an OpenAI Frontier Capability Group with more than 25,000 professionals to help enterprises use Codex and ChatGPT for software development and application modernization. Fanatics worked with OpenAI on Shop the Look, which recommends in stock team merchandise, and says views of recommended items rose 10%. Shared limits · Cognizant · Fanatics

OpenAI built Pando, a Git compatible backend for its monorepo, to handle the concurrent pushes, fetches and clones driven by Codex. OpenAI says median push latency fell by nearly five times and Pando now serves nearly one petabyte of Git data per day. OpenAI also uses activation classifiers, or probes, that read a model's internal activations in real time to flag unsafe or misaligned activity, stop responses, block tool calls and send messages for further review. OpenAI says the system runs on internal and external traffic for its most capable models, starting with GPT-5.6. Pando · Probes

MCP connectors and platforms

Bloomberg Law connected federal and state court dockets, filings and litigation intelligence to Claude through an MCP connector. LinkSquares launched a connector in Claude's directory for searching contracts and creating agreements from approved templates, and Mary Technology announced a Claude connector for investigating case facts. Everlaw unveiled its Omni agentic suite and connects its platform to Claude, ChatGPT, Copilot, Gemini, Harvey and Legora through MCP connectors. McGraw Hill will offer its course content in Microsoft 365 Copilot through an MCP plugin, with a limited beta in early 2027. Bloomberg Law · LinkSquares · Mary Technology · Everlaw · McGraw Hill

Litera connected its Lito legal AI to Box through the Model Context Protocol, so legal teams can search, retrieve and summarize Box documents in Lito while existing Box permissions and administrator controls stay in place. Litera and Box

Through athenahealth's new athenaPatient MCP, Claude Pro and Max users in the US can connect their electronic health records and ask about lab results, medications, conditions and visits. Patients authenticate with athenahealth, choose what to share and can revoke access. athenahealth

MSP360 launched an MCP server for managing backups in Claude and ChatGPT, free for Managed Backup customers and without a paid AI subscription. MSP360

Developer tools funding and acquisitions

&Truth, a public benefit corporation based in Palo Alto, came out of stealth with $70 million in seed funding for a system that uses formal verification and AI to prove whether code is safe, secure and correct. Crux Labs raised over $6 million in a round led by Benchstrength and Misneach to test AI agents across production traffic. Deepgram acquired Daily, the developer of the open source Pipecat framework for real time voice and video, and says Pipecat will remain open source. Freshworks acquired Robusta Dev and plans to add its AI outage investigation and root cause analysis to Freshservice. Neither deal's value was disclosed. &Truth · Crux Labs · Deepgram · Freshworks

CrowdStrike agreed to acquire agent observability company Fiddler AI and security testing startup Dreadnode for about $495 million combined, as reported, and will sell Fiddler's technology as Falcon Observe. CrowdStrike

Also

The Document Foundation says LibreOffice contributors may use LLM tools if they disclose that use, understand the code and pass human review; its plan to keep AI out of the default installation applies to built in user features. LibreOffice

Linus Torvalds released Linux 7.4-rc1 after a merge window of almost 20,000 commits, against the usual 15,000, and says it includes many fixes found or suggested by AI; Linux 7.4 may become the next long term support release. Linux 7.4-rc1

A Quickbase and Pave survey of 5,177 US employees and 527 leaders found that 24% of employees had built an app they use for work with AI; security reviews, integration, testing and limited support often kept those apps from wider use. Quickbase survey

Gartner's first Magic Quadrant for AI Agent Management Platforms places Microsoft, IBM and ServiceNow in the Leader quadrant. It evaluates tools that discover, govern, secure and monitor AI agents and track their costs and business value, and also assesses AWS, Salesforce and OneReach.ai. Gartner