blog
Blog — page 33 of 39
Posts 769–792 of 928, newest first.
Running npm install on untrusted code in a microVM
npm install is arbitrary code execution with a friendly progress bar. A malicious postinstall can read your SSH keys and tokens the second you install. Here's how to make that safe.
Firecracker Doesn't Use vhost-net (On Purpose)
QEMU reaches for vhost-net to hit line rate. Firecracker deliberately doesn't. Here's the full packet path — guest virtio-net to userspace device model to TAP — and why keeping the network device out of the kernel is a security decision, not an oversight.
Ajay KumarPer-Match Game Server Isolation on MicroVMs
Minecraft-style, Agar.io-style, and session-based match servers where each lobby gets its own microVM — untrusted mods contained, warm worlds snapshotted for instant join, torn down when empty.
Ajay KumarBest Secure Code Execution APIs in 2026
The honest roundup of hosted APIs for running untrusted, LLM-generated code — POST code, get stdout: PandaStack, E2B, Modal, Daytona, Vercel Sandbox, Fly.io, plus gVisor/Kata/WASM building blocks — judged by decision criteria, not a leaderboard.
Ajay KumarFirecracker vs Weave Ignite: VMM vs a Docker-UX runner
Ignite wasn't a rival to Firecracker — it was a management layer on top of it: image-to-VM, an ignited daemon, CNI networking, GitOps. Weaveworks is gone and the project is dormant, but the pattern it demonstrated is exactly how every production Firecracker platform is built.
Ajay KumarCopy-on-Write Memory: Why Forking a VM's RAM Is Cheap
Forking a VM means cloning gigabytes of RAM — which sounds impossibly slow until you realize nobody copies the bytes. Copy-on-write memory maps the snapshot's RAM MAP_PRIVATE so restored guests page-fault lazily and share every unwritten page. A clone is O(metadata), and pages diverge only when written. Here's the mechanism, and why it's what makes best-of-N agent branching cheap.
Ajay KumarSandboxing AI-Generated Terraform and Infrastructure-as-Code
An agent that writes Terraform and runs terraform apply can delete your prod database or exfiltrate your cloud keys — IaC tooling executes arbitrary providers and reads the environment. Run plan in a disposable microVM, gate apply behind a human review of the diff.
Running Long-Lived AI Agent Tasks in microVMs Without Paying for Idle
A deep-research or coding agent runs for minutes to hours but spends most of that wall-clock waiting on a token stream. Give it one durable sandbox that survives every step — and hibernate it while it thinks so you're not renting a hot CPU to do nothing.
Ajay KumarPer-Tenant Email Sending Isolation with microVMs
Your customers write email templates and send-time transforms. Running that in your shared worker is how one tenant's infinite Handlebars loop melts everyone's Tuesday. Give each send its own microVM.
Ajay KumarRunning AI Agent Eval Harnesses in Isolated microVMs
Every eval task needs a clean, reproducible box the agent can trash. Bake the task environment once, fork per attempt, run hundreds in parallel, and never let task 7 poison task 8.
Ajay KumarRunning Customer UDFs Safely in a SaaS with microVMs
The moment your product lets a customer write code — a workflow step, a validation hook, a transform — you're hosting untrusted code in your infra. Here's how to run it without betting the company on it.
Ajay KumarTesting LLM-Generated Database Migrations Safely in a Sandbox
An AI agent writing a migration is one hallucinated DROP TABLE away from ruining your night. Apply it to a disposable Postgres VM first, diff the schema, and throw the VM away.
Ajay KumarBuilding an Online Judge on microVMs
Contestants submit fork bombs, /proc snoopers, and code that tries to read the answer key. A microVM-per-submission judge makes that their problem, not your host's.
Ajay KumarHow Firecracker Schedules vCPUs: The Threading Model
A Firecracker vCPU is just a host thread wearing a convincing hat. Here's how those threads map to physical cores, how Linux CFS schedules them, and why you can pack far more vCPUs than cores.
Ajay KumarDebugging a Firecracker microVM That Won't Boot
A microVM that boots to a silent black hole is Firecracker's way of saying you assumed the rootfs, didn't you. Turn on the serial console and let the kernel tell you what actually went wrong.
Ajay KumarPandaStack vs Beam Cloud: which for AI code?
Beam Cloud is a serverless GPU/Python cloud for workloads you own; PandaStack runs untrusted agent code in per-task Firecracker microVMs. An honest comparison.
Ajay KumarThe Zygote Pattern: Fork One Warm Snapshot Into Thousands of MicroVMs
Bake one microVM with the heavy framework already imported, then fork it a thousand times. The children share the parent's pages copy-on-write, so N VMs cost far less than N× the RAM.
Ajay KumarMulti-Tenant Notebooks: microVM Sandboxes vs JupyterHub
JupyterHub is a great auth + proxy + spawner framework — but its isolation is only as strong as the spawner you pick. Here's where a microVM-per-user changes the math, and where it doesn't.
Ajay KumarSandboxing an AI Email-Triage Agent
An email-triage agent opens attachments, follows links, and does whatever the message body tells it to. Detonate each email in its own throwaway microVM, then delete it.
Ajay KumarRunning User-Uploaded Automation Scripts Safely
The moment your SaaS lets users write a "run this Python" step, you're running arbitrary strangers' code on your servers. Here's how to give each uploaded script its own throwaway microVM instead of a shared worker.
Ajay KumarPer-Tenant Isolation for RAG and Vector Search
The scariest bug in multi-tenant RAG is the one where tenant A's question retrieves tenant B's board deck. Here's why a shared index makes it inevitable, and how a per-tenant microVM plus a dedicated pgvector database turns the tenant boundary into hardware instead of an if-statement.
Ajay KumarFirecracker vs Microsoft Hyperlight, honestly compared
Hyperlight throws away the guest kernel to hit sub-millisecond starts for tiny WASM functions; Firecracker keeps a whole Linux guest so it can run anything. Same KVM lineage, opposite bets. Here's the fair comparison.
Ajay KumarFirecracker Snapshot/Restore vs CRIU Checkpoint/Restore
CRIU freezes a process tree — FDs, sockets, memory — back onto a shared host kernel. Firecracker freezes the whole microVM — guest kernel included — and just doesn't tell the guest anything happened. Same idea, different layer, very different gotchas.
Ajay KumarThe Serverless Cold-Start Problem, Explained
A cold start isn't one thing — it's a stack of them: provision a sandbox, init a runtime, load dependencies, run app init, wire up the network. This is where the milliseconds actually go, a taxonomy of the standard fixes and what each one really costs, and why snapshot-restore beats warm pools economically — plus the gotchas nobody warns you about.
Ajay Kumar