Blog — page 11 of 38
Running npm install in a MicroVM: Dependency Installation Is Arbitrary Code Execution
A postinstall script is an unsigned contract you agreed to by typing eight characters. Here is how to run dependency installation in a disposable microVM — and how snapshot forking makes it cheaper than what you do today.
Rendering user-authored templates safely: SSTI and the microVM fix
You shipped a text formatter so customers could edit their own emails. Depending on the engine, you may also have shipped them a REPL on your application server.
The Custom Code Step: Isolating User Code in a Low-Code Builder
The "Run JavaScript" block is where a no-code product admits it is a code product with a very good editor. Here is how to run it without betting the platform on it.
How to profile code running in a sandbox
Attaching a profiler to a pod is muscle memory. Inside a microVM, half of that muscle memory works and the other half quietly wastes your afternoon.
How to build a multi-tenant audit log your customers' auditors will accept
Every enterprise deal asks for an audit log. Most teams bolt one on in a sprint, discover it is really just their application logs wearing a suit, and then get to build it twice.
SLOs and error budgets for a code-execution platform
The standard SRE examples assume a long-lived request-serving service. A platform that boots machines on demand breaks all of them. Here's the version that survives contact with a sandbox fleet.
How to cut your sandbox compute bill: an engineer's playbook
The per-second rate is almost never why your ephemeral compute bill is high. Idle time, forgotten resources and repeated setup work are. Here is the order I attack them in.
Docker-in-Docker vs microVMs for CI builds
Your CI runners are containers and your jobs need to build images. Every answer to that trades privilege for convenience — here's the honest accounting of all four.
PandaStack vs Coder
These two products are shopped against each other constantly and compete almost never. One gives a person a machine for the week; the other gives a program a machine for four seconds. Naming that split is most of the decision.
The best self-hosted FaaS platforms in 2026
"Self-hosted serverless" is a phrase that has to fight itself — you didn't eliminate the servers, you adopted them. Here's the honest field of open-source FaaS platforms, and the container-isolation question most of these lists skip.
A Model File Is a Program: Per-Tenant Serving Isolation
You let customers upload their own models and you serve inference for them. Congratulations: you built a remote code execution endpoint with a friendly onboarding flow. Here's the real threat model and the isolation shape that survives it.
Your Tenant Wrote a Regex. The Regex Is the Malware.
Your customer didn't ship malware, they shipped a pattern with a nested quantifier — and one long stack trace turns it into a fleet-wide outage. The failure modes of tenant-supplied log parsing, and the validation path that catches them before production does.
Your Support Agent Has a Shell and Your Admin Token
The ticket body is untrusted input from a stranger, and you are feeding it to a model that has a shell and your admin API token. Here is the threat model and the per-ticket microVM shape that contains it.
Your Scale-to-Zero App Never Sleeps, and Bots Are Why
An app set to sleep after 15 idle minutes billed like an always-on server for a month. Nobody was using it — except an uptime monitor, four scanners, and our own health check. Here's the two-predicate classifier that fixed it.
Lease vs Heartbeat: How a Fleet Decides a Node Is Dead
You cannot tell a dead node from a slow network — you can only choose which way to be wrong. Heartbeats, leases, ownership columns, and the negative-caching bug that 503s a perfectly healthy fleet.
PID namespaces and process isolation, explained: what CLONE_NEWPID actually gives you
A PID namespace gives you a private set of process numbers and a process the kernel treats as init. It is excellent plumbing and it is not a security boundary, and almost every bug you will hit lives in the gap between those two sentences.
Container Escape CVEs, by Class: What the Pattern Tells You
Container escapes aren't a list of unlucky bugs — they're four recurring shapes, plus one category that will never get a CVE because it's working as designed.
tmpfs in a microVM: The Filesystem That Eats Your RAM
Writing a gigabyte into /tmp inside a microVM does not consume disk. It consumes the guest RAM your process was counting on — and if you snapshot that VM, the gigabyte gets cloned into every copy.
PandaStack vs Google Cloud Run: Honest Head-to-Head
Cloud Run runs your container and owns its lifecycle. PandaStack hands you a microVM and lets you hold it. Most teams should use Cloud Run — this is about the narrow set who shouldn't.
Northflank vs Railway: Control or Momentum?
Both let you ship a repo without operating a cluster, and both are good at it. The choice is really about how much control you want versus how fast you want to be productive — plus the lock-in question everyone skips.
How a Sandbox Scheduler Decides Where Your VM Lands
Every sandbox platform has a scheduler picking which machine your VM lands on. Here's the scoring function we actually run, the staleness window that made five simultaneous creates pile onto one host, and why the fix wasn't smarter scoring.
How Many Sandboxes Fit on One Machine?
Everyone wants a number. The real answer depends on which resource binds first — and on most microVM fleets, the host memory metric you'd naturally scale on is actively lying to you.
Where Scale-to-Zero Wake Time Actually Goes
Scale-to-zero sounds like a hypervisor problem. When we instrumented a real wake, the VM restore was 1.4 seconds of 14 — the rest was our own orchestration doing careful things in the wrong order.
How to Vet a Code Execution Vendor's Security
If your AI agent runs model-generated code, you've outsourced a security boundary. Here are the questions worth asking a vendor, why SOC 2 answers almost none of them, and what the honest answers sound like.