Security

"Dangerously" stops being dangerous — and you control exactly how safe.

Every agent runs in an isolated DevBox with a security profile you choose — from open experimentation to air-gapped, read-only, no-secrets compliance.

The Security Problem

Running AI agents unattended shouldn’t mean disabling all security.

Today, the only way to run AI coding agents uninterrupted is to disable the protections that exist for a good reason.

Evidence

Don't take our word for it. Take theirs.

The all-access modes that let an agent run uninterrupted? Every maker tells you to keep them inside an isolated environment.

Claude Code--dangerously-skip-permissions

Only use this mode in isolated environments like containers, VMs, or dev containers without internet access, where Claude Code cannot damage your host system.

Permission modes docs
OpenAI Codex--dangerously-bypass-approvals-and-sandbox · --yolo

Run every command without approvals or sandboxing. Only use inside an externally hardened environment.

Codex CLI reference
GitHub Copilot--yolo · --allow-all

It is strongly recommended that you only use these options in an isolated environment.

Copilot CLI — allowing tools

And the everyday auto modes they recommend instead are still — by their own admission — not a safety guarantee.

Claude Code · auto mode

Auto mode is a research preview. It reduces prompts but does not guarantee safety.

Permission modes docs

OpenAI Codex · Auto · auto-review

Auto-review improves the default operating point for long-running agentic work, but it is not a deterministic security guarantee. It should complement, not replace, good sandbox design, monitoring, and organization-specific policy.

Codex auto-review docs

Anthropic said the same about auto mode publicly — keep it in a devbox / isolated environment. @claudeai on X

A DevBox is that isolated environment — disposable, off your laptop, away from your credentials. So you can run the modes their makers reserve for a sandbox.

Honesty note: Isolation reduces risk, it doesn’t erase it — even a DevBox can leak credentials to a malicious repo. We say isolated + disposable, never safe.

Real Risks

Destructive command execution

rm -rf, git push --force, system config changes — no approval required.

Secret leakage

Env vars, API keys, and tokens accessible to prompt injection attacks.

Lateral movement

Access to VPNs, cloud services, and internal networks from a compromised agent.

Environment contamination

Unwanted dependency changes, repo state corruption affecting your whole team.

No isolation between tasks

Multiple agents share the same filesystem, credentials, and permissions.

Isolation, by default

Isolated environments where autonomy is safe.

Instead of running agents with dangerous permissions on your machine, ClusterCode spins up isolated DevBoxes where each agent works unrestricted — safely.

Runs in its own DevBox

Isolated from the host, from other agents, and from your credentials.

Bypass flags are safe

The DevBox is disposable and sandboxed — nothing escapes.

Truly uninterrupted sessions

No confirmations, no interruptions. The agent works until it’s done.

Scoped access only

Cloned repo, dependencies, and the env vars you configure. Nothing more.

Real parallelism

5, 10, 15 simultaneous DevBoxes — each on a different ticket.

Before

$ agent --dangerously-skip-all-checks

Full access to your machine, your secrets, your network.

With ClusterCode

DevBox #1 — ticket-92
DevBox #2 — ticket-93
DevBox #3 — ticket-94

Each agent isolated. Bypass flags are safe. Nothing escapes.

Security Profiles

From experimentation to air-gapped compliance — one toggle.

Open
Network
Full internet
Filesystem
Read-write
Secrets / env vars
Injected
Hardening
default
Runaway-process protection*
Generous4 GB RAM · CPU/PIDs uncapped
Best for
Trusted dev
StandardDefault
Network
Dev registries only
Filesystem
Read-write
Secrets / env vars
Injected
Hardening
cap-drop ALL, no-new-privileges
Runaway-process protection*
Tuned2 CPU · 4 GB · 256 procs
Best for
Everyday work
Strict
Network
Your allowlist only
Filesystem
Read-only + carve-outs
Secrets / env vars
Injected
Hardening
cap-drop ALL, no-new-privileges
Runaway-process protection*
Lean1 CPU · 2 GB · 128 procs
Best for
Sensitive repos
Offline
Network
No internet
Filesystem
Read-only + carve-outs
Secrets / env vars
Not injected
Hardening
cap-drop ALL, no-new-privileges
Runaway-process protection*
Lean1 CPU · 2 GB · 128 procs
Best for
Compliance / air-gapped

* Hard ceilings on RAM (every profile) plus CPU cores and process count (Standard and up), enforced by the kernel — one runaway or malicious agent can't exhaust the machine (fork bombs, memory blow-ups) or starve its neighbors.

Network Egress Control

Allowed ✓

github.comregistry.npmjs.orgapi.anthropic.compypi.org+ your custom allowlist
DevBox
DNS allowlist
filter

Blocked ✕

evil.compaste-my-secrets.iorandom-exfil.netunknown-api.xyzeverything not on the list

Your AI agent can reach npm. It cannot reach the open internet.

Non-allowlisted hostnames return NXDOMAIN — code, secrets, and test data have nowhere to go.

How ClusterCode Compares

ClusterCode
Isolated environment
Per-task DevBox
Network egress control
Configurable allowlist
Read-only filesystem option
Yes
Withhold secrets entirely
Offline profile
Self-hosted / your hardware
Yes
Setup
Out of the box — one toggle
Your own containers (DIY)
Isolated environment
Yes — you build it
Network egress control
DIY — you wire & maintain
Read-only filesystem option
Manual setup
Withhold secrets entirely
Manual
Self-hosted / your hardware
Yes
Setup
Build & maintain it all yourself
Cloud agents (black-box)
Isolated environment
Vendor-hosted
Network egress control
Not user-configurable
Read-only filesystem option
Not exposed
Withhold secrets entirely
Not exposed
Self-hosted / your hardware
No
Setup
Vendor-controlled
Local CLI agents (no isolation)
Isolated environment
Runs on your machine
Network egress control
None
Read-only filesystem option
No
Withhold secrets entirely
Runs with your env
Self-hosted / your hardware
N/A
Setup
No controls to set up

Comparison reflects publicly documented capabilities as of mid-2026; offerings evolve. You can replicate much of this with your own containers — ClusterCode just ships it built-in, enforced, and one toggle per launch.

Trust & compliance

How we run the platform itself.

The controls above protect each agent run. These are the operational practices that protect the orchestrator, your account, and the limited data we do hold.

Encryption in transit and at rest

Traffic to ClusterCode is served over TLS. The account, billing, and telemetry data we store are encrypted at rest.

Least-privilege access

Scoped access controls and continuous monitoring govern who and what can reach our production systems.

Your code never reaches us

By design, your source code, agent outputs, and secrets stay on your own hardware — shrinking what is ever at risk on our side.

Incident response

If a breach affects you, we notify you without undue delay and notify the relevant supervisory authority within 72 hours where the law requires it.

Vetted sub-processors

Stripe, Clerk, and Microsoft Azure (US hosting and transactional email) — each under contract that restricts how it may use your data.

Honest compliance posture

ClusterCode is not yet SOC 2 certified, and the Service is not intended for HIPAA, PCI DSS, or other regulated-data workloads.

Found a vulnerability?

We welcome responsible disclosure. Email security@clustercode.io — our policy and contact details are also published at /.well-known/security.txt.

Every DevBox has a security profile — you choose the level.

Network egress controls prevent code and data exfiltration.

Read-only filesystem on Strict and Offline; capability hardening on Standard and up.

Offline profile injects no secrets and has zero network — for air-gapped compliance.

With ClusterCode, "Dangerously" stops being dangerous — and you control exactly how safe.