blog

Blog — page 26 of 39

Posts 601–624 of 928, newest first.

9 min read

Letting an AI Agent Run Integration Tests Against Real APIs

Unit tests are safe to run anywhere. Integration tests need real credentials and real egress — exactly the two capabilities you least want to hand a model on a shared host. Here's the shape that works.

ai-agentstestingmicrovmAjay Kumar
9 min read

Running User-Generated Game Mods in Isolated microVMs

A mod is arbitrary code from a stranger that you execute on your infrastructure. "We removed the io library" is not a security boundary — a guest kernel is.

gamingmoddingmicrovmAjay Kumar
8 min read

Per-Tenant Workflow Workers in Isolated microVMs

A shared worker pool runs every tenant's activity code in one process on one kernel. One runaway retry loop starves everyone, and "we run customer code in a thread with a timeout" is not a security model.

workflowsmulti-tenantmicrovmAjay Kumar
9 min read

Robotics Simulation and RL Rollouts in Isolated microVMs

Simulation is embarrassingly parallel and pathologically environment-sensitive. A baked microVM snapshot is the pinned sim environment — and a per-sandbox network namespace finally shuts ROS 2 discovery up.

roboticssimulationreinforcement-learningAjay Kumar
11 min read

Running IoT and Embedded Firmware Emulation in Disposable MicroVMs

A firmware image is a filesystem, an init, and a stranger's opinion about what should happen at boot. You are going to run all three. Do it somewhere you can delete.

firmware-analysisiot-securityemulationAjay Kumar
10 min read

Isolating Per-Tenant Geospatial Processing Jobs in MicroVMs

A customer uploads a shapefile and your worker decides, based on those bytes, which of GDAL's hundred-plus drivers to run. That's not a data pipeline — that's attacker-controlled parser dispatch with a database credential in the room.

geospatialgdalmulti-tenantAjay Kumar
10 min read

Sandboxing AI-Agent 3D Rendering and Asset Pipelines in MicroVMs

Your agent writes a Blender script and something has to run it. Meanwhile the scene file it was handed can execute Python the moment it opens. Give that job a machine you can throw away.

3d-renderingblenderasset-pipelineAjay Kumar
10 min read

Processing PHI in Per-Job microVMs: Isolation for Regulated Healthcare Data

The senders are hospitals with 1998-era interface engines, the formats are containers full of compressed pixel data and XML, and the parsers are C. Give every PHI job a machine you can afford to lose — and know exactly which parts of compliance that does and does not buy you.

phihealthcarehipaaAjay Kumar
11 min read

Firecracker vs Xen: two generations of the same idea

Xen taught the industry what "cloud virtualization" meant. Firecracker came out of the same company with a deliberately narrower answer. This is not old vs new — it's two architectures tuned for different eras.

firecrackerxenhypervisorAjay Kumar
10 min read

Nested virtualization and Firecracker: what actually works

"Can I do nested virtualization with Firecracker?" is two completely different questions wearing one trench coat. One has a yes-with-caveats answer. The other is a flat no, and that's usually fine.

nested-virtualizationfirecrackerkvmAjay Kumar
10 min read

What Runs as PID 1 Inside a MicroVM (and Why It Matters)

The kernel boots, mounts a rootfs, and executes exactly one program. That program's job description is short, strange, and easy to get wrong — which is why sandboxes hang, leak zombies, lose their last log line, or take thirty seconds to stop.

firecrackermicrovminitAjay Kumar
10 min read

virtio-blk Discard and TRIM in Firecracker, Explained

A job downloads 6 GB, does its work, deletes it, and the host-side image is still 6 GB. Deleting a file is a metadata edit in the guest's own allocator; the host never hears about it unless somebody issues a discard. Every link in that chain fails silently.

firecrackervirtio-blkdiscardAjay Kumar
11 min read

Landlock explained: unprivileged filesystem sandboxing in the Linux kernel, and where it stops

Landlock is the rare Linux security module that needs no root and no sysadmin: a process hands the kernel a list of directories it promises never to leave, and the kernel holds it to that promise forever. Excellent inside a boundary. Not a boundary.

landlocklinux-kernellsmAjay Kumar
11 min read

The Best Ephemeral CI Runner Platforms in 2026

A shared CI runner is a build cache, a credential store, and a mutable filesystem that every pull request gets to write to. Ephemeral means one job, one fresh machine, destroyed after — the interesting question is what that costs you.

ci-cdephemeral-runnersgithub-actionsAjay Kumar
11 min read

Running Fuzzing Harnesses and Crash Reproduction in MicroVMs

A fuzzer is a professional vandal you hired on purpose. Its entire job is to push your parser into states the author never imagined — so don't run it on a kernel you share with anything you care about.

fuzzingsecuritymicroVMsAjay Kumar
10 min read

Isolating Per-Tenant CSV and Bulk Import Pipelines in MicroVMs

Your "Import your data" button accepts arbitrary bytes from anyone with a trial account. Zip bombs, 4GB single lines, formula injection, and a spreadsheet parser made of C. Give it a machine you can throw away.

csv-importbulk-importmulti-tenantAjay Kumar
11 min read

Database Branching with Copy-on-Write MicroVMs

Branching a database means giving someone a private, writable, instantly-available copy of real data. There are two ways to build it: copy-on-write at the storage page level, or forking the entire machine — disk and memory — so the branch inherits the warm buffer pool and a postmaster that never noticed it was cloned.

database-branchingcopy-on-writepostgresAjay Kumar
11 min read

Isolating Smart Contract Simulation and Forked-Chain Testing in MicroVMs

Simulating a contract means running two kinds of untrusted code: the bytecode, and the ordinary JavaScript that deploys it. One of those can read your archive-node API key out of the environment. Give each simulation its own machine, exactly one upstream, and a pinned block height.

smart-contractsevmsimulationAjay Kumar
11 min read

Firecracker Hugepages for Guest Memory, Explained

Backing guest RAM with 2 MiB hugetlbfs pages means one page fault covers 2 MiB instead of 4 KiB — 512x fewer faults to populate the same memory on restore. The part nobody documents: hugepage-ness is a property of the snapshot, not a runtime flag, and Firecracker will only restore such a snapshot through the userfaultfd backend.

firecrackerhugepagesmemoryAjay Kumar
11 min read

What's Actually Inside a Firecracker Snapshot

One file is kilobytes of structured VMM state. The other is a flat, sparse copy of every byte of guest RAM. Almost every clever thing anyone has ever done with microVM snapshots is a trick played on the second file.

firecrackersnapshotsinternalsAjay Kumar
10 min read

Firecracker Metrics and Logger FIFOs, Explained

Firecracker has no /metrics endpoint on purpose — it pushes JSON to a file descriptor your supervisor owns. Which means your collector is now load-bearing infrastructure, and a named pipe nobody is reading can wedge a running microVM on a write(). Here's the design, the footgun, and the collector loop that survives it.

firecrackerinternalsobservabilityAjay Kumar
11 min read

Firecracker vs AWS Nitro Enclaves: Two Different Threat Models

One protects your host from other people's code. The other protects data from your own operators — including you. People reach for the wrong one constantly, and only find out when they try to SSH into an enclave that has no network, no disk, and no shell.

firecrackernitro-enclavesconfidential-computingAjay Kumar
11 min read

PandaStack vs GitHub Codespaces

Both give you a computer in the cloud with your repo in it. One is optimised for a person with an editor; the other for a program spawning hundreds of VMs an hour. The tell is whether a for loop is creating them.

comparisondev-environmentscodespacesAjay Kumar
11 min read

Best Sandbox APIs for TypeScript AI Agents in 2026

You're building an agent in TypeScript — Vercel AI SDK, LangChain.js, Mastra, or a hand-rolled tool loop — and something has to run the code the model wrote. This is about the SDK you'll actually live in: typed results, streaming, AbortSignal, ESM, bundle size, plus an honest pass over the field.

comparisontypescriptai-agentsAjay Kumar