Edera

If you're interested in Kubernetes, the Linux kernel, AI agents, or anything eBPF-related, you won't want to miss this session 😊 Our upcoming webinar will address the various security considerations (and possible shortfalls) that infrastructure teams will need to understand before choosing an agent sandbox.

An AI agent waiting on a human shouldn’t cost the same as one doing work. In a lot of sandbox setups, it does. Compute stays reserved while the agent sits idle, and every new task or RL run pays to build the same environment from scratch. Project Lunchbox changes that math on the Kubernetes clusters you already run. Snapshot a running environment and suspend it while the agent waits. Resume it when it’s needed. Or prepare an environment once and fork it. Each copy is an isolated microVM that shares memory with the original until it changes a page. We can fork thousands of them very quickly. Making that work meant operating on a different clock than Kubernetes. As our CTO Alex Zenla told LinuxInsider: “At the point we are freezing and CoW, we are talking milliseconds, whereas Kubernetes frequently measures things in seconds.” When state can be shared, tasks take less time. That’s where the cost comes down. Jack M. Germain has the full story: https://lnkd.in/gmV3FrDR

The best question from our last webinar didn't get a full answer. It's getting a full hour. "How does Edera compare to the open source projects like Kubernetes Agent Sandbox and others like Kata and gVisor?" The honest answer depends on what you're running, where you're running it, and what failure you're designing against. So instead of answering it ourselves, we brought in outside voices. Jon Ceanfaglione (Sparrow Labs) and Timothy St. Clair (Raikiri) join Alex Zenla to work through it: -Where the kernel boundary actually sits in each model -Why Agent Sandbox is an API layer rather than a competing runtime -The production tradeoffs: syscall overhead, eBPF visibility, GPU isolation, and nested virtualization Join us and bring more questions! https://lnkd.in/g9K4b_kC

Running AI agents at scale comes with a fair amount of infrastructure pain that doesn't get talked about enough. Isolation is one part of it. Then there's provisioning, networking, storage, resource management, environment setup, debugging, upgrades, and keeping the whole thing running reliably. Whether you're using Kata, gVisor, Firecracker, containers, or building your own isolation stack, someone has to own all of this. And that engineering time comes out of somewhere, usually the product roadmap. I'm a big fan of open source isolation technologies. If you have the time, resources, money, AI, and a roadmap where you can keep throwing more engineering and infrastructure at the problem, you can build a really good agent execution platform with these technologies. But you're also signing up to maintain it. AI can help you build and automate much of it, but it doesn't eliminate the operational burden. For teams without that luxury, how much engineering capacity are you willing to spend maintaining the infrastructure instead of building what your customers actually need? This is why I like what we've done with Edera's agent isolation ROI calculator. It attempts to put some numbers around engineering capacity, infrastructure spend, environment initialization, execution overhead, and isolation costs. These are estimates, not guaranteed savings, and the economics depend on your workloads and existing infrastructure. If building and maintaining your own isolation platform gives you a competitive advantage, absolutely do it. Otherwise, it's worth understanding how much of your engineering roadmap you're spending just keeping the infrastructure underneath your agents working. https://lnkd.in/e5iNdEDa

Well, this is a fun way to start a Monday! 😁 Edera made it into The Register! And apparently, our take on KVM security was described as “none-too-gentle.” I'll take that as a compliment. :) The article covers PwnEscape (CVE-2026-99999), a critical guest-to-host escape involving Linux networking code. The vulnerability sits in the Linux networking stack, but its impact can cross the VM boundary. Guest isolation depends on more than the hypervisor code alone. At Edera, we're building on both Xen and KVM. Each comes with different architectural tradeoffs and attack surfaces. Our job is to understand those boundaries, isolate workloads, and limit what an attacker can reach if something goes wrong. No hypervisor is bug-free. What matters is how far an attacker can get when something breaks. Pretty cool to see this discussion getting attention in The Register! https://lnkd.in/ed2szHB4

CVE-2026-99999 is the latest in a long list of Linux kernel vulnerabilities surfacing this year. As more businesses continue to look for viable agent sandboxing solutions, the choice between Xen or KVM for virtualisation matters more than ever. In this latest Edera blog post, I explored some of the security benefits and potential shortcomings of either approach. https://lnkd.in/d6vCs52Q

48% of organizations using AI coding assistants say more security issues are reaching production since they adopted them. Detection tools were trained on human-written code. They have no baseline for code written by agents. As agents write and run more code, more of it goes unexamined. IDC study's take: “Containment doesn't ask enterprises to verify it first; it bounds what any of it can affect, using infrastructure most enterprises have already deployed at scale.” . Read the analyst brief from IDC https://lnkd.in/g-A7an4Z

If you're planning on being at KubeCon Salt Lake City, we will be bringing the same in-demand interactive runtime demos and fun CTF challenges. Make sure to swing by booth 226 to compete.

Edera has been popping up at events all over the place in recent weeks! Just today, I was speaking at Cloud Native Dublin about container and agent workload isolation through the lens of Webernetes. If you’ve missed us at any of these events, we do our best to share recaps of what we’ve been up to, the talks we’ve been giving, and our thoughts on the evolving landscape of cloud-native sandboxes. We even run the odd CTF challenge when we’re in town! Last week, we were at Open Source Summit Europe in Prague, Czechia. Here’s the condensed recap :) 🔗 https://lnkd.in/d5B_zktC

"We take security seriously" is not a warranty. Neither is a well-curated SOC 2 report. Especially when, as we learned this year, some of them can apparently be delivered faster than you can say "observation period." Your SaaS vendors share clusters, kernels and identity systems across every customer they have. When one tenant gets breached, you could be next. And what you have to go on is a questionnaire they filled out about themselves and an audit report that may have been written from a template. I wrote down the 5 warranties that we at Edera believe that every CISO should demand in writing, dated, and checkable later: 1️⃣ Tenant isolation at every layer, down to the kernel 2️⃣ Phishing-resistant MFA and zero standing privilege 3️⃣ Supply chain accountability for every subprocessor 4️⃣ Workload identity and access that actually expires 5️⃣ Tamper-evident audit logs you can pull via API The sad truth is that each one of these warranties is backed by a real breach, where they weren't asked for. Snowflake. Okta. CircleCI. A container escape that could reach ~100 pods. Read it, bring this list to your next QBR or renewal, and then forward it to a peer. One CISO asking alone is easy to defer. A hundred asking together isn't.👇 https://lnkd.in/g6BPjJFA