If you're choosing how to run code you don't trust, WebAssembly and Firecracker both belong on the shortlist — and they solve the problem from opposite ends. WASM starts from a program that can do nothing and hands it capabilities one by one; the module can only touch what you explicitly imported. Firecracker starts from a full Linux machine that can do everything, and wraps a hardware wall around it so that "everything" stays inside one guest. WASM asks "may I?" before every syscall; a container just does it and hopes; a microVM lets the guest do it and contains the blast radius. This post is the untrusted-code comparison specifically: what each isolation model actually guarantees, what WASM genuinely cannot run, and the honest rule for picking one.
The one-line verdict, up front: WASM wins when you control compilation to a constrained target and want maximum density plus microsecond-class startup. Firecracker (and platforms built on it, like PandaStack) wins when you must run arbitrary, unmodified, full-OS workloads — the kind an LLM generates when it decides at runtime to pip install something with a native extension. Everything below is why.
Two isolation models, one question
WASM's model is capability-based software isolation. A WebAssembly module runs in a sandbox defined by the language and runtime, not by a kernel. Its linear memory is a single bounds-checked region — every load and store is range-checked, so a module physically cannot address memory outside its own heap. The callstack is structurally inaccessible (return addresses aren't values the module can forge), and every control transfer lands on a type-checked target. Critically, there are no ambient syscalls: the module reaches the outside world only through host functions you import. WASI is the standardized version of that host interface — a capability-gated set of imports for files, sockets, clocks, and the like. Hand a module a single pre-opened directory and that's all the filesystem it will ever see. Hand it nothing and it's pure compute in a box.
Firecracker's model is hardware isolation. Each microVM boots a real Linux kernel and is confined by KVM — the CPU's virtualization extensions enforce the boundary, not a syscall filter in software. The guest talks to the host through a deliberately tiny device model (virtio-net, virtio-block, virtio-vsock, a serial console — almost nothing else to attack), and Firecracker layers a jailer (chroot + dropped privileges) and a default-on seccomp filter around the VMM process. The guest can run any Linux binary in any language, unmodified, because there's a genuine OS underneath. The cost is a larger memory footprint and millisecond-not-microsecond startup: you're booting or restoring a machine, not instantiating a function.
Side by side
- Isolation boundary — WASM: software fault isolation + capability set at the runtime level (no guest kernel). Firecracker: hardware virtualization (KVM) with a full guest kernel.
- What it runs — WASM: only code compiled to the wasm target, within imported capabilities. Firecracker: any unmodified Linux binary, any language, plus subprocesses, package managers, and compilers.
- Startup — WASM: fast, microsecond-to-millisecond-class instantiation (verify against your runtime's docs). Firecracker: milliseconds — ~3s first cold boot, or ~49ms when restoring a snapshot.
- Memory footprint — WASM: kilobytes to low megabytes per instance. Firecracker: a few MB of VMM overhead plus the guest's baked RAM (whole-OS footprint).
- Syscalls / subprocesses — WASM: no ambient syscalls; no fork/exec, so no arbitrary subprocesses. Firecracker: full POSIX — fork, exec, threads, shells, everything.
- Attack surface you own — WASM: the host runtime's compiler/JIT and the capabilities you grant. Firecracker: the VMM device model + hypervisor (small, heavily audited).
- Best fit — WASM: code you compile and constrain, at high density. Firecracker: arbitrary, unpredictable, full-OS workloads (e.g. AI-generated code).
What WASM can't run (the part that decides it)
For untrusted code the choice usually comes down to a binary question: does your workload compile to wasm and live within a capability set, or not? WASM is more capable than the stale lore suggests — as of WASI Preview 2, modules can do capability-gated filesystem access, TCP/UDP sockets, and HTTP, so "WASM can't do networking" is a retired Preview 1 complaint. But three hard limits still decide the arbitrary-code case in 2026:
- No fork/exec. WASI doesn't implement POSIX process management. A module can't spawn a subprocess or shell out — which rules out anything that runs another program (a build step, a linter,
git, a headless browser). - No arbitrary native binaries. A wasm runtime executes wasm modules, not
/usr/bin/ffmpeg,gcc,apt, orpsql. If it wasn't compiled to wasm, it doesn't run — barring research-grade emulation like container2wasm, which isn't a default path. - Native extensions only if pre-ported. NumPy and pandas work in Pyodide because the project specifically compiled those C/Fortran extensions to wasm. An arbitrary
pip installof a package with a native extension fails unless a wasm build already exists — and you can't predict which package a user or an agent will reach for.
There's also a subtler cost people forget: with WASM, the host runtime is now part of your trusted computing base. The whole boundary rests on the runtime's compiler/JIT correctly enforcing bounds checks and control-flow integrity in software. That's a real, actively-maintained surface — and because the untrusted module and the runtime share one address space, a speculative side-channel (Spectre-class) attack has attacker and victim in the same process. Wasmtime's own docs treat Spectre mitigation as partial and ongoing. A hardware VM boundary is generally considered stronger against that class because attacker and victim sit in separate address spaces — stronger, not immune. KVM has had real guest-to-host escape CVEs too. Neither is magic; state the difference as "stronger," never "safe."
Where WASM genuinely wins
Credit where it's due: for the workloads it fits, WASM is dramatically lighter and faster to start than any VM, and pretending otherwise would be dishonest. Instantiation is microsecond-to-millisecond-class with per-instance memory in the kilobytes-to-low-megabytes range — one to two orders of magnitude lighter than a microVM. (Exact numbers are runtime- and module-dependent; check your runtime's docs rather than trusting a single canonical figure.) That's why edge platforms run untrusted per-request functions as wasm at densities a VM model can't touch. Reach for WASM when:
- You control the source — the code is, or can be, compiled to wasm, and you constrain the language and dependencies.
- You want extreme density and near-instant cold starts: edge functions, plugin systems, per-request isolation, untrusted expressions or user-supplied transforms evaluated at scale.
- The workload lives entirely inside a capability set — no subprocesses, no arbitrary native binaries, no full POSIX environment needed.
- The untrusted surface is something you can tightly scope: a rules engine, a math expression, a scripting extension to your own host application.
In that situation WASM isn't a compromise — it's the better tool, and a hardware VM would be over-paying for a boundary you can get cheaper while losing the density that made the workload economical. A minimal WASI run looks like this — a single module, granted exactly one directory:
# Compile Rust to the WASI target, then run it under wasmtime.
# The module gets ONE pre-opened dir (--dir) and nothing else —
# no ambient filesystem, no network, no subprocesses.
cargo build --target wasm32-wasip1 --release
wasmtime run --dir=./data::/data target/wasm32-wasip1/release/transform.wasm
# Try to shell out or read /etc/passwd from inside that module?
# There's no syscall to do it — the capability was never imported.Where Firecracker wins
Reach for a microVM the moment the workload needs the OS, the boundary, or the unpredictability handled:
- You must run arbitrary, unmodified code and native binaries: a real shell,
pip installof anything, compilers, databases, multiple processes, package managers, a full filesystem, an actual kernel. - You're running mutually-untrusting multi-tenant workloads and want a hardware boundary rather than a shared in-process one.
- You can't predict the dependencies in advance — the defining property of AI-generated code, where the model decides at runtime what to install and execute.
- You need whole-machine snapshot and fork — freeze an environment mid-task and branch it — which is a VM-level capability, not a function-level one.
This is exactly the AI-agent and general code-execution case, and it's worth being blunt about why. An agent's entire value is that it runs commands you didn't write and couldn't fully anticipate: it shells out, installs native-dependency packages, spawns subprocesses, compiles things, hits the network. Every one of those is something WASM either can't do or can only do if the dependency was pre-ported. The honest reason to pick Firecracker here isn't that it's "more secure" in the abstract — it's that it's often the only model that can run the workload at all, and it happens to arrive with a hardware boundary attached.
Where PandaStack lands
PandaStack runs on Firecracker because that's where arbitrary, untrusted, model-generated code belongs. Every sandbox is a Firecracker microVM with its own guest kernel (Ubuntu 24.04 guest), isolated by KVM — not a shared-kernel container, and not an in-process wasm sandbox. So a pip install of anything, a real shell, subprocesses, compilers, and the full long tail of native-dependency Python all just work, because there's a genuine Linux kernel underneath rather than a capability set you had to anticipate:
from pandastack import Sandbox
sbx = Sandbox.create(template="code-interpreter")
# Arbitrary, unmodified native code — no wasm build required.
# In WASM this would fail the moment a native extension appears.
sbx.exec("pip install duckdb pandas")
out = sbx.exec("python -c \"import duckdb; print(duckdb.sql('select 42').fetchone())\"")
print(out.stdout) # (42,)
# Shell out, spawn subprocesses, compile — all fine on a real kernel.
sbx.exec("gcc --version && git init /tmp/repo")The classic objection to the VM model is cost, and that's the part the engineering answers. There's no warm pool of idle VMs; every create restores a baked snapshot on demand — ~49ms in the fast path, 179ms p50, about 203ms p99, versus ~3s only on the first cold boot before a snapshot exists. Same-host forks branch a running environment via copy-on-write memory plus an XFS-reflink rootfs. The core is open-source and self-hostable on your own Linux KVM hosts. None of that makes Firecracker the right answer for a tightly-scoped plugin — if your workload fits WASM, use WASM. It makes Firecracker the right answer for the workload it's actually for: arbitrary code you didn't write and can't fully predict.