Ericsson Software Technology had a fork problem. Like a lot of large engineering organizations, it kept private copies of open source projects, patched them in-house and carried those patches forward release after release. Then the EU Cyber Resilience Act (CRA) made that habit expensive. To meet its CRA obligations, Ericsson eliminated its private forks and pushed more than 1,400 dependency updates and security fixes back upstream.
That case study, released by the Open Source Security Foundation (OpenSSF) this week, is a useful lens for almost everything announced around Open Source Summit Europe in Prague, October 7-9. The CRA's mandatory vulnerability and incident reporting requirements took effect in September. AI is making bugs easier to find, which shrinks the time teams have to fix them. And AI agents are now writing, running and shipping code with permissions nobody fully reviewed.
The Linux Foundation's news package for the event includes more than two dozen announcements. Taken together, they point developers toward four practical shifts: fix things upstream, enforce agent permissions below the model, treat compliance artifacts like build outputs and design for portability from the start.
The fork tax now has a number
The fifth annual State of Open Source in Europe report, from Linux Foundation Europe, LF Research and NeoNephos, puts a price on private forks. Forty-eight percent of organizations maintain them. Those that do spend an average of 311 hours per release cycle patching 9.5 forks. And 28% can't track those costs at all.
Regulation is pushing teams to look harder at that number. Sixty-three percent of organizations have increased investment in open source dependency management because of regulatory pressure. Yet 31% still lack a formal open source governance framework. Organizations with advanced governance report an average return of more than 4x on their open source investment, compared with 3.6x for those without formal governance.
"Open source gives organizations a way to reduce dependence on any single vendor or geography, but consuming open source alone doesn't give you control over the technology," said Thierry Carrez, general manager of Linux Foundation Europe. "By contributing upstream and participating in open governance, European organizations earn a seat at the global table making greater technology autonomy possible."
OpenSSF is trying to make the upstream path easier to follow. Alongside the Ericsson case study, the foundation published a CRA Readiness practitioner's guide that turns policy analysis into steps for maintainers and vendors, plus a CRA Readiness User Journey and role-based guidance paths for developers, security engineers and OSPO leaders. It also shipped OpenBao v2.6, the open source secrets manager, with per-namespace sealing and a new workflow engine for cross-plugin communication. And it accepted BOMHort, a Kubernetes-native SBOM visualization and governance tool, into the OpenSSF Sandbox.
JetBrains joined OpenSSF as a General Member, along with A-Team Systems, Emphere and DACHS IT. "AI is changing how software is built and creating new security challenges, making it more important than ever that developers can understand, verify and trust the software they produce," said Katherine Druckman, head of community and partnership engagement at JetBrains.
The financial sector is taking the same shared-fix approach. The Fintech Open Source Foundation (FINOS) announced that the Open Source Enterprise Resiliency Alliance (OSERA) is now operational, with initial funding from seven Premier members, including founding members Citi, Deutsche Bank, Morgan Stanley and Royal Bank of Canada. Banks run many of the same open source stacks. OSERA gives them a shared, transparent way to identify, fix and verify vulnerabilities in that software instead of each bank patching the same bug on its own.
Agents need limits the model can't talk its way around
AI agents are moving into production faster than the governance around them. The 2026 State of OSPOs and Open Source Management report from LF Research and the TODO Group found that 69% of organizations are prototyping or running agentic tools for open source program office workflows, and 18% already run agentic AI in production. The Europe report found that 94% of organizations are using or piloting generative AI coding tools.
Several announcements this week share one design principle: enforce what an agent can do at the infrastructure layer, not in the prompt.
The clearest example is nono, an open source runtime from Luke Hinds, who created Sigstore while at Red Hat. nono puts AI agents, MCP servers and LLM workloads in a lightweight, kernel-enforced sandbox. When an agent tries to open a file, call a tool or reach the network, nono checks the action against policies written in code. The operating system kernel enforces the decision, not the agent. If the policy says no, the agent can't do it, regardless of what the model decides. nono also keeps an independent audit trail of agent actions.
Since launching in March, nono has logged 328,278 installs, 4,249 GitHub stars and more than 100 maintainers. It's Apache 2.0 licensed, with commercial support from Hinds' company, nolabs.
"One of the lessons from Sigstore was that developers will embrace security when it makes a hard problem simple and stays out of their way," Hinds said. "Developers should not have to choose between moving fast with AI and keeping systems safe."
James Carnegie, staff security engineer at Datadog, said nono is "the only sandbox that gives us both fine-grained, per-command policies and sophisticated credential management that fits existing, complex real-world toolchains."
Coder is applying the same idea to cloud coding agents. Coder Agent Relay, with SpaceXAI as launch partner, lets Cursor Cloud Agents run inside Coder workspaces on infrastructure the customer already operates. Cursor still runs the agent loop, including inference and planning. But tool calls execute in Coder environments on the customer's network, so source code, secrets and internal services stay on machines the customer controls. Each agent environment is sandboxed, ephemeral and scoped to a single task. Coder says a prompt injection that pushes an agent toward unauthorized resources gets blocked at the environment layer rather than left for the model to refuse. Every run produces a log of what the agent accessed, executed, changed and was blocked from doing. The integration is in private preview with design partners.
"Enterprises never rejected AI agents. They rejected the deployment model," said Rob Whiteley, CEO of Coder. "What has kept this work out of production is everything that comes after the pilot: where the code runs, who has access, what the agent can reach, and what record exists afterward. Those are infrastructure questions."
At the industry level, the Open Secure AI Alliance, started by NVIDIA, moved to the Linux Foundation in September. Its goal is an open defensive stack covering models, inference, agents, context, identity, policy and enforcement that organizations can inspect and run on their own infrastructure. Its first project, the Shared AI Findings Exchange (SAFE), is designed to collect AI security incident findings confidentially, notify affected parties and turn recurring failures into evidence-based controls.
"AI security improves when organizations openly share what they learn," said Justin Boitano, vice president of enterprise AI at NVIDIA.
Even infrastructure tooling is building agent access with guardrails from the start. Meshery, the cloud native management plane for Kubernetes, became a CNCF incubating project this week. Its roadmap includes a Meshery MCP Server that gives AI assistants read-only access to the Meshery Registry and cluster state. Meshery is now the fifth highest-velocity project in the CNCF, with a 350% increase in code commits over the past year, more than 7,000 contributors and nearly 12,000 GitHub stars.
Compliance is turning into a build output
SBOMs used to be paperwork. Regulation is turning them into an engineering deliverable, and the industry is trying to agree on what a good one looks like.
The Linux Foundation released the OpenChain Automotive SBOM Framework 1.0, an open framework for how automakers and suppliers create and exchange SBOMs. It defines a structure for describing components and dependencies, required and recommended data fields with completeness and quality criteria, and usage scenarios for exchanging SBOMs between suppliers and OEMs. It builds on SPDX (ISO/IEC 5962) rather than replacing it. Toyota, Hitachi Solutions, Nissan and Volvo Cars all backed the release, and the framework is on GitHub.
"Effective use of SBOMs is more than just generating or publishing them on-demand," said Tim Mackey, head of software supply chain risk strategy at Black Duck. "Differing requirements and expectations between organizations and regulators open the door to SBOM quality issues – something that a shared framework and clear guidance on structures and data quality helps resolve at scale."
The Civil Infrastructure Platform (CIP), now 10 years old, achieved compliance with IEC 62443-4-1 and 4-2, the international cybersecurity standards for industrial automation and control systems. CIP estimates that companies building on its processes and reusable artifacts can cut their own IEC 62443 compliance effort by up to 70%. A new LF Research report on CIP's first decade explains how the project turns standard Linux LTS kernels into 10-year Super Long-Term Support (SLTS) baselines. That matters for anyone shipping long-lived industrial hardware, especially with CRA penalties that can reach €15 million or 2.5% of global annual turnover.
The ELISA Project, which focuses on safety-critical Linux, brought BASIL, Cregit and stress-ng into its ecosystem, adding tools for code analysis, testing, system reliability and performance.
On the distro side, AlmaLinux rewrote its certification suite. Hardware certification now takes under 10 minutes for most systems. It installs with dnf install alma-certify and can run from live media without installing AlmaLinux or disrupting existing workloads. The foundation also added software certification, so vendors and users can list products and supported releases at catalog.almalinux.org. AlmaLinux is clear that this is "a statement about compatibility," not a regulatory or security certification.
Embedded turns 10, the kernel turns 35
The Zephyr Project marked its 10th anniversary, with Texas Instruments and Infineon Technologies upgrading to Platinum membership.
Embedd launched Zephyr Project Assembler to attack one of Zephyr's hardest problems. Zephyr runs on more than 1,000 boards, but a production product's custom board isn't one of them. Teams still hand-write devicetree sources and overlays, wire up Kconfig and work through build failures before writing any application code. Embedd's agentic system builds a complete Zephyr project from a PCB design, a bill of materials or a plain description of the device. Its agents work from a digital twin of the MCU built from vendor reference manuals, so pin assignments come from the silicon rather than a model's guess. Every project is validated with west build before hand-off. Developers can edit the configuration visually or directly in VS Code. It's generally available as a VS Code extension and on the web.
The Linux kernel turns 35 this year, and an LF Research report shows the project still growing. The kernel sees nearly 10 changes every hour, with a new release every eight to nine weeks. The 7.0 through 7.2 releases set records for first-time contributors. The report also points to built-in real-time support (PREEMPT_RT), mainlined in kernel 6.12 after a 20-year effort, and to AI tooling now helping maintainers with patch review and security auditing.
openKylin 3.0 takes AI further into the operating system. The release moves to a Linux 7.0 kernel base, upgrades more than 180 core components, adds compatibility with the RISC-V RVA23 standard and supports agent tools and frameworks, including Claude Code and LangChain, instead of tying users to a single AI provider.
Sovereignty becomes an architecture decision
Digital sovereignty was the loudest theme in Prague, and the data backs it up. In the Europe report, 94% of organizations call digital sovereignty important, and 61% call it very important. A separate OpenSearch Software Foundation and LF Research study found that 83% of organizations run or plan to run AI workloads, 78% consider cross-vendor interoperability critical when choosing data platforms, and 71% say deploying independently of a single vendor or cloud provider is strategically important.
For developers, sovereignty shows up as portability. Can the workload move? Can the team change providers without a rewrite?
OVHcloud upgraded to Platinum membership in the Cloud Native Computing Foundation (CNCF). Its Managed Kubernetes Service runs thousands of production clusters on Kubernetes and Cilium across 20 public cloud regions, and its engineers contributed the Kubernetes Operator for Harbor. "Independence in technology does not mean building everything alone," said Octave Klaba, CEO and founder of OVHcloud. "It means having the freedom to build on open standards, contribute to the technologies you rely on, and avoid being locked into a single ecosystem."
The NeoNephos Foundation, which works toward an open European cloud ecosystem, now has 21 members after Giant Swarm and SVA joined. It's opening more of its work to individual contributors, including a Hack the Garden hackathon around Gardener, its Kubernetes cluster management project. Nextcloud Hub 26 Summer brings the open source Euro-Office suite to the desktop, and Nextcloud says more than half of the release's improvements to Calendar, Mail and Contacts came from community contributors.
AWS also became a Platinum Member of the Linux Foundation. It now holds memberships in 25 Linux Foundation projects and foundations. "When we build infrastructure that solves a problem beyond AWS, we look for a neutral home to help sustain it," said Stormy Peters, head of open source at AWS.
One more to watch: OpenGrid
Not every announcement fits the developer toolchain, but OpenGrid is worth a look for anyone who works with data or APIs. More than 2,500 gigawatts of renewable, storage and large-load projects are stuck in grid connection queues worldwide, according to the International Energy Agency (IEA). Part of the bottleneck is software: grid planners work with fragmented data and incompatible tools.
The Linux Foundation's new OpenGrid initiative isn't building a replacement planning tool. It's building the shared layer underneath: open standards and schemas, programmatic APIs and a Data Hub for validated, model-ready datasets. Its first application, the OpenGrid Translator, is due in 2027 and will move data and models between commercial platforms and open source tools. Breakthrough Energy GRIDS and Google.org are among the backers.
"Today, too many plans rely on data others cannot verify and tools much of the world cannot afford," said Alice Yake, vice president at Breakthrough Energy GRIDS and chair of the OpenGrid Management Committee.
What this means for your team
The common thread in Prague is that enforcement is moving down the stack. Agent permissions are enforced in the kernel or the workspace, not in a system prompt. Compliance is enforced by a shared framework and a build step, not a spreadsheet someone fills out before an audit. And vendor independence is designed into the architecture instead of negotiated in the contract.
For developers, that's mostly good news. It means less guesswork and fewer late-stage surprises. But it also means the work shifts left. If your team keeps private forks, someone should be counting the hours. If you're putting coding agents into production, you need a clear answer to what each one can touch and who answers for it when it gets something wrong. My view hasn't changed: whoever builds and deploys the agent owns what it does. Tools like nono and Agent Relay make that ownership something you can enforce, not just promise.
Ericsson didn't upstream 1,400 fixes because it was generous. It did it because carrying them alone stopped making sense. Many more teams are about to do that same math.