Plain-English posts on running coding agents in isolated boxes: safety, parallel sessions, snapshots and forks, verification on a real desktop, and how Sessionboxer compares with Devin, Cursor, Codex, Claude Code and OpenHands.
Claude Code 2.1.328 starts Opus 4.6 and Sonnet 4.6 sessions on Bedrock, Agent Platform and Foundry in auto mode. Who it affects and which setting to change.
Dragos seeded two million prompts with fake keys and caught three hacked AI hosts and prompt-logging resellers. What a coding agent sends, and what to check.
Cursor Cloud Agents run Cursor's agent in VMs on Cursor's side, or on a worker you register. How that compares with Cursor in a box on your own machine.
JetBrains fixed four Junie CLI flaws that let the agent write outside your project or run commands you never approved. Which build to run, and what to set.
Coding agents write good Mermaid, but a terminal only prints the source. Where to view the diagram, which viewer to avoid for private code, and how to check it.
Lumia showed that whoever controls a TLS-inspecting proxy can replace Claude Code, Cursor or VS Code updates with malware. What it needs, and what to change.
Copilot cloud agent works on GitHub's machines and pushes to a copilot/ branch. How that compares with running Copilot and other agents in boxes you host.
Microsoft showed that one approved GitHub device code can steer a Copilot CLI session, plant memory and read private repos. How it works, and what to check.
Codex Cloud runs Codex tasks in VMs on OpenAI's side, from environments Codex sets up. How that differs from running Codex in a box on your own machine.
Coder now runs its own coding agent in its control plane, on your API keys. How that differs from Sessionboxer's box per session, and who should pick which.
Darktrace steered Claude Code through a Claude memory store and pulled a 1 GB file out via the Files API. What it needed, and what to check on your machines.
Claude Code and Codex both let you type while the agent works, but a queued message reaches the model at different times. What each does, and when to use it.
Since November 10, stolen GitHub tokens have pushed a secret stealer into thousands of repositories. How to check yours, and what to change for your agents.
Claude Squad runs each agent in a tmux session and a git worktree, as your user. Sessionboxer puts each one in a Docker container. What each does, and costs.
Claude Code 2.1.325 no longer runs sandboxed Bash commands without a prompt. What changed, which runs it stops, and when to set autoAllowBashIfSandboxed.
Depi turned one GitHub issue into a path to Codex npm releases through an agent that labels issues. How the chain worked, and what to change in your workflows.
Dagger's container-use runs your agent's commands in a container on a git branch. Sessionboxer runs the agent itself in the box. What each one isolates.
An AI security agent left a researcher's sandbox and pulled customer data from a WorkOS tool. What WorkOS found, and how to limit what your agents can reach.
Kimi CLI is archived and Kimi Code CLI replaces it. How to put the new CLI in a container, log in or pass a key without a prompt, and pick a safe mode.
Claude Code can now run cloud sessions on runners in your network. What stays on your machines, what you must operate, and how it compares to Sessionboxer.
Claude Code 2.1.321 loads mods Claude writes without asking in auto and bypass modes, and their tool calls skip ask rules. What changed, and how to stop it.
Claude Code 2.1.319 can deny calls that would wait on a permission prompt and keep the turn going. What it covers, how it differs from dontAsk, where to set it.
Mistral Vibe in a container needs a key without a prompt and a run that stops on its own. How to pass the key, pick a profile and cap turns, cost and time.
Glow Labs found coding agents installing archived MCP servers from memory, one with an unpatched RCE. How to find them, what to use instead, how to block them.
OpenHands Agent Canvas and Sessionboxer both run coding agents on your own hardware. What each one isolates, which agents they run, and who fits which.
Apiiro found 1,122 public copies of Claude Code's settings.local.json holding a secret. How to check yours, what to rotate, and how to stop it happening again.
Claude Code runs well in a codespace until the idle timer or the 12-hour limit stops it. How to set it up, keep the login, and what the agent can reach.
Conductor Cloud goes read-only on November 13 and starts deleting data on November 20. How to move workspaces to your Mac, export, and plan after that.
An agent in Docker Sandboxes could plant a symlink so your next sbx cp read or overwrote a file outside the sandbox. Version 0.51.0 fixes it. What to check.
GitHub CLI 2.99 can attach a video to a pull request from the terminal, so a coding agent can show its UI change. How it works, the limits, and who can see it.
A researcher showed that one line in a cloned repo's AGENTS.md makes Claude Code's auto mode approve nearly every command. What to check and set today.
Qwen OAuth is gone, so Qwen Code in a container needs an API key. How to pass it in, which approval mode to pick, and why its own sandbox does not fit here.
Forever Security showed that one page or local HTML file could run code or read your whole disk through five AI coding tools' built-in browsers. What to check.
Conductor runs parallel agents in worktrees on your Mac or in its Vercel cloud. Sessionboxer runs each one in a container on a machine you own. How they differ.
JFrog found 42 GitHub repositories with hooks that start a worm when a coding agent opens the folder. Which files to look for, and how to clean up safely.
T3 Code and Sessionboxer both drive Claude Code, Codex and Cursor with your own subscription. One runs them on your machine, the other in a box per session.
For eleven hours on October 29, Cursor's marketplace served a fake Claude Code extension that planted an SSH key. How to check, how to clean up, what to change.
Cursor CLI in a container needs a login without a browser, and its own sandbox fails under Docker's defaults. How to set up both, and which flags to pass.
Claude Code 2.1.314 adds a setting that makes Write and Edit refuse paths outside your project in every mode. What it covers, what it leaves to the shell.
OpenCode lets its agent run shell commands without asking, so a container is a good place for it. How to build one, bring in your login and set the permissions.
A sandboxed command could swap a folder for a symlink and make Claude Code's Edit tool write outside the project on macOS. What it says and what to check.
Copilot CLI in a container has no keychain and no browser, and it reads any GitHub token in its environment. How to log it in, and where to approve every tool.
Cursor's agent can now click and type in any app on your computer. What it sees on a shared desktop, what the docs ask you to grant, and where else it can run.
Your company's gateway setup runs an apiKeyHelper on your laptop, and a container cannot run it. How to send Claude Code in Docker through the gateway anyway.
Amp's self-hosted orbs run the agent's tools on compute you provide, and Amp keeps the agent loop. What stays in your network, what a pause keeps, the price.
Claude Code Review, Copilot and Codex review pull requests on their own servers. What each one does, and how to run the same review on a machine you own.
Gemini CLI in a container stops at folder trust and at Google login. What each check does, how to pass it on purpose, and where approving every tool is safe.
Claude Code 2.1.304 and later run slow MCP tools as background tasks that can stop to ask you a question. What happens in headless runs, and what to set.
Claude Code, Cursor, Gemini CLI and Codex each undo part of what an agent did. Which part, what shell commands leave behind, and how to get the rest back.
Claude Code 2.1.309 stops shell startup files from running outside the Bash sandbox and protects the --mcp-config file. What changed, and what to check.
Claude Code's /voice needs a local microphone and a claude.ai login, so it stops in a container or over SSH. Why, and three ways to dictate to a remote agent.
Claude Code 2.1.305 added sandbox.allowNestedUserNamespaces so Chromium and rootless podman can run in its Linux sandbox. It also lets them undo write denies.
Codex CLI in a container needs a login without a browser and a choice between its own sandbox and Docker's. How to set up both, and what each one leaves open.
Pillar Security showed that a remote MCP server could run code through OpenCode's OAuth login on Windows and WSL. Which versions, the fix, and what to do now.
CERT-In's October 15 advisory says AI agents read company data with credentials left in config files and public repos. What it asks for, and how to apply it.
Claude Code 2.1.302 starts some claude -p runs in Manual mode instead of auto and turns on forked subagents in scripts. What changed, and which flags to set.
Subagents report back, agent teams talk to each other, separate sessions get their own machine. What each one shares, what it costs, and when to use it.
OpenCode says a breach hit 1.4% of Console users who had linked provider keys. How to tell if yours were there, what to revoke, and how to cap the next bill.
A coding agent can change more files in ten minutes than you can read in an hour. An order for reviewing its diff, the tools that help, and what undo misses.
Antigravity CLI now approves safe commands by itself and sends sandbox escapes to a second agent. What that covers, what it does not, and what to set first.
Your coding agent pushed with your personal GitHub account, or got a 403. How git and gh pick an account, and how to bind each repository to the right one.
Since October 13, GitHub blocks pushes that contain two new Anthropic key types. Where the block applies, where it does not, and what to tell your agent.
Copilot CLI, Claude Code, Codex and Devin can now fix a pull request's failing checks and review comments. How each one works, and what to check first.
Claude Code 2.1.299 summarizes a 200K-token conversation left idle for 54 minutes, before its prompt cache expires. When that helps and how to turn it off.
Most coding agent integrations assume GitHub. How to give an agent your self-hosted Bitbucket: the token, the git login, pull requests and their builds.
GitHub now records the account that pushed each commit. If your coding agent pushes with your login, that account is you. Give the agent one of its own.
Most coding agents run in Linux containers, so they cannot build your .exe or click through it. Your three options on Windows, and what each one costs.
Inside a container, adb cannot see the phone on your desk. The usual Docker flags, why they break or overshare, and how to hand one phone to one agent.
Before Claude Code 2.1.295, a policy hook that crashed or timed out let the action run. What onFailure "block" does, where it does nothing, and how to test it.
When Claude Code says you've hit your session limit, the task stops until the reset. What the limits are, what you can do, and how to resume the turn by itself.
Copilot CLI's /model now lists models from a running Ollama. A local model alone is not offline mode. What changed, how to set it up, and what still leaves.
Behind Zscaler or Cloudflare WARP, an agent in a container fails with "self signed certificate in certificate chain". Why, and how to fix it without -k.
An agent that runs docker compose usually gets your Docker socket or a privileged container. Both can reach your machine. The options, and what each costs.
VS Code 1.141 can confine Copilot's terminal commands to the files and networks you allow. It is off by default, the network stays open, and it is not a VM.
Claude Code can run on a timer with /loop, Desktop tasks, cloud Routines or GitHub Actions. Where each one runs, what it can reach, and when a run is skipped.
Six ways to run a coding agent in its own machine: where it runs, whose account pays, whether you can watch the screen, snapshot the box or switch agents.
Claude Code's docs say to skip permission prompts only inside a container or VM. What the flag turns off, what still protects you, and how to set up that box.
In April 2026 a Cursor agent wiped a startup's database and backups during a staging task. The lesson is not more prompts. It is what the agent can reach.
Claude Code 2.1.287 added mods: JavaScript hooks that rewrite prompts and tool calls and run with your permissions. That changes where the agent should run.
Solo developer, tech lead, QA engineer, on-call SRE, student or manager: the right coding agent setup is different for each. A plain guide to picking yours.
Devin 2.2 and Cursor Cloud Agents test their changes with computer use. Get the same loop on your own machine, after every turn, with a captioned recording.
Each Claude Code turn sends far more than your message. See the system prompt, tool schemas, MCP servers and history that fill the context window, and the cost.