all postsblog / topic / code execution

blog · topic

Code execution sandboxes

Executing code you did not write — from an LLM, a user, or a pull request — is a sandboxing problem first and a runtime problem second. These posts cover code interpreter architecture, Python and notebook execution, timeouts and resource limits, and why the isolation boundary matters more than the language runtime.

230 posts

9 min read

What Is Vibe Coding? (And Where the Generated Code Should Actually Run)

Vibe coding is prompting a model for code, not reading it, and steering by whether the thing works. That's fine for a weekend toy. It is less fine when the unread code runs npm install on the laptop that holds your SSH keys.

ai-agentsconceptssandboxAjay Kumar
10 min read

How to Create an AI Agent Sandbox (Step by Step, 2026)

Isolation boundary, disposable filesystem, network policy, three timeouts, guaranteed teardown, and one run_code tool the model calls. Three ways to build it, with code for each.

ai-agentssandboxtutorialAjay Kumar
11 min read

Checking Untrusted Proofs (Lean, Coq/Rocq, Isabelle) in Isolated microVMs

Proof checking looks like the safest workload there is, right up until you remember that Lean elaborates a file by running it. An RL pipeline generating millions of candidate proofs has two problems: what a proof can do to your machines, and what it can make your checker say.

microvmfirecrackertheorem-provingAjay Kumar
10 min read

LLM-as-Judge vs Execution-Based Evals: Why Your Code Evals Need a Sandbox

A judge model will call wrong code idiomatic. Running it tells the truth, if the code can't edit the tests, exit early, or patch the assertion. Execute first, judge the remainder.

ai-agentsevalsllm-as-judgeAjay Kumar
9 min read

Daytona vs Vercel Sandbox (2026): Which One for AI Code Execution?

Daytona grew up as a dev-environment platform; Vercel Sandbox grew up next to a deploy platform. Both now run code your model wrote. The question is whether your agent needs a workspace or a box shaped like a request.

comparisondaytonavercelAjay Kumar
10 min read

Analysing APKs Nobody on Your Team Wrote: One MicroVM Per Sample

The thing that breaks your mobile-app scanning pipeline is not a shell. It is a jadx run that pins every core on one crafted DEX, an apktool that writes small files until the volume runs out of inodes, and a zip entry named ../../etc/cron.d/x. Here is the isolation shape that survives that, and what it still cannot do.

mobile-securityapk-analysismicrovmAjay Kumar
11 min read

PandaStack vs Kubernetes Jobs: Where "Just Create a Job" Stops Being the Right Answer

Your first instinct is to create a Kubernetes Job, and that instinct is usually correct. Here is the hardened Job you should actually write, the four mechanisms where the model breaks for untrusted per-request code, and the shim you will own if you push through anyway.

kuberneteskubernetes-jobsmulti-tenancyAjay Kumar
10 min read

The Best Sandbox APIs for PHP and Laravel AI Agents in 2026

PHP cannot sandbox PHP — eval is not a boundary and disable_functions is a denylist over a vast standard library. Here is the honest field of options for running model-written PHP, and the PHP-shaped questions that decide which one fits.

comparisonphplaravelAjay Kumar
11 min read

Every In-Process Python Sandbox Leaks, and the Reason Is Structural

Removing a name from a namespace removes a name, not reachability. CPython's object graph has no partitions, and ctypes is a documented feature whose purpose is calling arbitrary native code. Here is where the boundary actually has to go.

pythonsandboxingrestrictedpythonAjay Kumar

More posts in this topic

221 more, newest first.