In this edition, covering 20 to 24 October:
- Moonshot AI released Kimi K3.1 Pro, an open weight model it says scores 65.7% on Terminal-Bench 4.0, close to Claude Opus 5.5 at 66.4%.
- Recursive released Flint, a 40 billion parameter open weight coding model it says was built largely by its own AI system.
- Blacksmith, a platform for building and testing code, raised $66 million in a Series C led by Sequoia at a $1 billion valuation.
New coding models
Moonshot AI introduced Kimi K3.1 Pro, a 3T class open weight model for coding and agent tasks with a 1 million token context, plus a Highspeed version it says runs two to three times faster. Moonshot says it scores 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5, and full weights are due by 4 November. US startup Recursive released Flint, a 40 billion parameter open weight coding model; CEO Richard Socher says its Eureka Machine system improved the model's training, architecture and chip specific code. Kimi K3.1 Pro · Flint
Security research on code and agents
Anthropic says Claude analyzed more than five million lines of code across 25 open source projects and found 493 issues, including in OpenSSH, whose maintainers fixed access to about 4,500 lines of unauthenticated code. Claude Security now scans repositories continuously in a public beta, and Anthropic and Theorem used Claude to produce formal security proofs for 123 OSS Fuzz libraries. 493 issues · Claude Security · Theorem
Anthropic Fellows tested Claude Code Auto Mode and OpenAI Codex Guardian against persistent malicious coding agents and found weaknesses involving prompt injection, multi agent workflows and unmonitored edits. Cloudflare uses OpenAI models through Daybreak to find vulnerabilities in customer code and propose patches, and says four developers rebuilt a product with Codex in three days for $1,100 in tokens. Hashgraph Online's open source Guard can block risky coding agent actions before they run. Coding agent monitors · Cloudflare · Guard
Testing and tooling
Blacksmith raised $66 million in a Series C led by Sequoia at a $1 billion valuation for its platform for building and testing code, including code written by AI. mabl introduced a testing harness for verifying code from coding agents before a pull request and was named a Challenger in Gartner's first Magic Quadrant for agentic software testing. Unstructured launched Transform models for parsing PDFs, scans and tables, and NTT developed a way to link documentation, code and tests so AI can maintain older systems. Blacksmith · mabl harness · Gartner · Unstructured · NTT
A Linux Foundation survey found generative AI cuts open source code writing time by 52%, while 25% of incoming pull requests are low quality. Google is piloting Gemini assistance in its software engineer coding interviews. Linux Foundation · Google interviews
MCP and agent platforms
Workday opened its agent platform to outside tools and models through MCP, Zoom introduced a Unified Context Layer that shares meeting, chat and document data with AI agents, and Microsoft released Azure MCP Server 3.0 with stricter authentication defaults. Cloudflare turned on its WebMCP beta across nearly 100,000 Commerce.com storefronts. The Agentic AI Foundation added Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. Workday · Zoom · Azure MCP · WebMCP · Agentic AI Foundation
New MCP servers came from Haystack, Capital Markets Gateway, Anrok and Financial Cents, and Harvey made its Connector Library generally available. Reduck raised $3 million for agents that use browser scripts on sites without APIs. Haystack · CMG · Anrok · Financial Cents · Harvey · Reduck
Companies
Tech Mahindra and SpaceXSI will offer Grok Build to help enterprises modernize applications. Replit appointed former AWS executive Michael Greenwald as president of corporate affairs as it expands its government business. Tech Mahindra · Replit