US cyber agency says "A-OK" to open source
Plus: Cloudflare's busy week in open source, a big upgrade for MCP,and much more.
Hi folks,
This week’s lead story looks at a notable piece of guidance from the US government’s cybersecurity agency that essentially tells federal agencies to default to open source software and treat vendor lock-in as a security risk.
Elsewhere, Cloudflare has had a busy week in the open source realm; MCP gets a major enterprise upgrade; Nscale doles out $1.65 billion for the company behind open source distributed computing framework Ray; and much more.
As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.
Paul
<Open issue>
CISA to federal agencies: open source is fine, vendor lock-in is the real risk
Open source software supply chain attacks grew in both volume and impact in 2025 and early 2026, according to a report published last week by Google’s Threat Intelligence Group and its cybersecurity group Mandiant. On the very same day, the US Cybersecurity and Infrastructure Security Agency (CISA) — the federal body responsible for protecting America’s critical digital infrastructure — published a 31-page document telling every US federal civilian agency how to think about open source software security.
None of what CISA says will surprise anyone already working in open source. The argument that open source software is verifiable in a way that proprietary software is, is not new. What is different is who is making it now, and under whose authority. This is a Trump administration document, published under two executive orders on federal cybersecurity.
Its central claim is thus:
“All software carries risk, and OSS is no more or less risky than other software. The key distinction is that, with OSS, agencies can directly assess code quality and security, rather than relying solely on vendor assurances.”
The document explicitly flags vendor lock-in as a security risk, noting that open standards and modifiable code protect agencies from “proprietary dependency traps.”
Moreover, the CISA document tells agencies to adopt a “default open source” approach to custom-developed software, requiring justification for keeping code closed rather than the other way around. To structure how agencies evaluate open source projects, it introduces the C4 framework, grouping trustworthiness into four areas: the health of the codebase, the strength and structure of the community behind it, the conduct of its governance — things like whether there is a vulnerability disclosure policy and mandatory code review — and the security of its default configuration.
On AI specifically, the guidance addresses something the open source community and broader technological world has been grappling with for some time: the difference between “open-weight” and “open source”. A permissive license is not sufficient, CISA says — agencies must have access to the full training data and training pipeline before treating a model as open source for risk management purposes. Without that, it should be treated as proprietary software with incomplete provenance.
Read more: CISA | Google/Mandiant
<Patch notes>
MCP gets its enterprise upgrade
The Model Context Protocol (MCP) — the standard for connecting AI agents to tools and data that Anthropic open-sourced in 2024 — has received its largest update since launch, adding stateless architecture, a formal deprecation policy, and hardened authentication to make it production-ready for enterprise deployments.
Read more: The Agentic AI Foundation (AAIF)
Cloudflare’s big open source week
In the space of around one week, Cloudflare open-sourced three projects: pvcli, a command-line debugger for privacy-preserving protocols such as Oblivious HTTP; an AI-powered triage tool that cleared Astro’s GitHub issue backlog from 200+ issues to near zero is also now available for one and all; and Cloudflare OS, the platform the company has been running internally since May that gives every employee an AI agent workspace built around company context, systems, and governance.
Read more: The New Stack | The New Stack | Cloudflare
A Ray of light for Nscale
London-based AI cloud provider Nscale has agreed to acquire Anyscale — the commercial steward of Ray, the open source distributed computing framework donated to the PyTorch Foundation in 2025 — for $1.65 billion. Ray itself stays community-governed and free to run anywhere; Nscale is buying the engineering team, commercial platform, and customers including Coinbase.
Read more: Nscale
Alibaba may start charging big users of Qwen
Reuters reports that Alibaba plans to require large commercial users of its next open-weight Qwen model to share revenue — following a similar move by rival Moonshot — underscoring a key distinction between open-weight AI systems and traditional open source software, which typically permits commercial use without any revenue-sharing requirements
Read more: Reuters
Eclipse and OWASP join forces on EU cyber compliance
The Eclipse Foundation and OWASP have partnered to help open source maintainers prepare for the EU’s Cyber Resilience Act, whose vulnerability reporting obligations take effect on 11 September, with joint guidance, readiness webinars, and SBOM adoption support.
Read more: Press Release
Codeberg votes to ban vibe-coded projects
Codeberg, the Forgejo-based open source hosting platform, has updated its terms of use to prohibit AI-generated “vibe-coded” projects following a member vote — arguing that LLM-generated, unmaintained repositories pollute the FLOSS commons and strain infrastructure with resource usage no human team could justify. The platform has also committed to never using user data to train LLMs.
Read more: Codeberg
JetBrains open-sources KotlinLLM
JetBrains has open-sourced KotlinLLM, a research prototype IntelliJ IDEA plugin that lets Kotlin developers write function calls whose bodies are generated at runtime by an LLM — rather than calling a model on every request or running a separate agent. Once generated, the code runs as plain Kotlin without any further model calls.
Read more: JetBrains
Microsoft open-sources a polyglot unit testing agent
Microsoft has open-sourced a unit-test generation agent that learns from a repository’s conventions, writes tests across multiple languages including Python, Go, and .NET, and verifies they build and pass before finishing — completing 92% of tasks in internal benchmarks versus 79% for stock GitHub Copilot.
Read more: Microsoft
<Final commit>
Google open-sources offline pocket translator
A creative team at Google has open-sourced a working universal translator built on a Raspberry Pi 5, running Google’s Gemma 4 model entirely on-device with no internet connection required after setup.
The device — housed in a custom 3D-printed case — uses a microphone to capture speech, transcribes it using the open source Moonshine model, translates it via Gemma, and speaks the result back through a speaker. Two people can face each other and speak in different languages, each hearing the other translated in near real time.
The 3D print files for the case are included in the public repo, so you can build your own.
Read more: Google Gemma (X) | GitHub