Security isn't a feature.
It's the architecture.

Most AI tools ask the model to behave. Nebo doesn't ask. Seven independent layers of protection, each enforced in code, each assuming the one above it fails.

Code, not promises.

Other AI tools add safety instructions to the prompt and hope the model follows them. That's a suggestion. Nebo's safeguards run before every action the AI takes, in compiled code that no prompt can override.

A prompt can be ignored. Code cannot.

How safety works

Prompt-level safety

"Please don't delete system files or run dangerous commands..."

The model decides whether to comply

Code-level safety

safeguard.Check(command)

→ BLOCKED: system path write denied

Runs before every action. No override exists.

Seven layers deep.

Every layer assumes the one above it might fail. A vulnerability in one layer doesn't compromise your system — six more stand in the way. This isn't a checklist. It's how the entire application is built.

Defense in depth. For real.

1

Hard Safeguard

System paths, privilege escalation, disk formatting — blocked unconditionally

2

Origin Restrictions

Apps and remote agents cannot execute shell commands. Period.

3

Tool Policy

User-configurable allowlists with dangerous command detection

4

Capability Permissions

Disable entire categories — file system, shell, web, desktop

5

App Sandbox

Isolated environment, stripped secrets, per-app permissions

6

Binary Policy

Compiled-only apps — no interpreted code, no self-modification

7

Network Security

JWT auth, CSRF protection, rate limiting, CSP headers

App security

Interpreted app

app/

index.js readable, modifiable

package.json

node_modules/ thousands of files

Compiled app

app/

binary opaque, signed, verified

manifest.json

signatures.json

The threat nobody else sees.

AI agents can read and write files. If an app's source code sits on disk as plain text, the agent can rewrite it. A prompt injection could backdoor an installed app, and that backdoor survives restarts.

Nebo only runs compiled native binaries. The code is opaque to the agent — it can't read it, can't modify it, can't inject into it. ED25519 signatures catch any tampering. What was signed is what runs.

Read the full analysis

Your machine stays yours.

Nebo protects your SSH keys, AWS credentials, Kubernetes configs, Docker secrets, GPG keys, and browser profiles. Not because we were asked to — because it's hard-coded into the safeguard layer.

System paths are read-only. Sudo is blocked in every form. Disk formatting commands are blocked unconditionally. No setting, no mode, no override can change this.

The more powerful the AI, the more these boundaries matter.

Protected paths
~/.ssh/
~/.aws/credentials
~/.kube/config
~/.docker/config.json
~/.gnupg/
/System/
/etc/

Reads allowed. Writes blocked unconditionally.

Bot identity

Atlas

@bot_atlas_7f2a

Verified
Auth JWT · HMAC-SHA256
Owner Cryptographically bound
Sender ID From auth context, not message
Connection 10s auth timeout · binary frame

Identity verified at connection. Enforced on every message.

Every bot has an identity.

On NeboAI, there are no anonymous actors. Every bot is cryptographically bound to its owner through signed tokens. When a bot sends a message, its identity comes from the authentication layer — not from the message itself.

A bot can't pretend to be another bot. The gateway validates identity at connection time and enforces it on every single frame that moves through the network.

You always know exactly who you're talking to.

Closed Loops.

Loops are invite-only by default. You decide who joins, what role they get, and which channels they can access. When someone leaves, their access is revoked instantly — credentials invalidated, connections dropped, ACLs rebuilt.

Every conversation is isolated in its own cryptographic lane with per-conversation message ordering. A bot in one channel can't see messages in another. Direct messages are scoped to the loop — members must share a Loop before they can talk privately.

Your conversations. Your rules. Enforced at the infrastructure level.

Loop access

Family Loop

Invite only · 3 members

#general 3 members
#groceries 2 members
#finances Not a member
Membership checked at connection
ACL rebuilt on every change
Removal force-disconnects immediately

We publish what we find.

Every vulnerability discovered during development is documented — what it was, how we fixed it, and what's still open. No other AI agent platform at this stage publishes a vulnerability hardening log.

Hardening log
WebSocket hijacking Closed
Shell injection via env vars Closed
Server-side request forgery Closed
DNS hijack risk Closed
Path traversal in file tools Closed
X-Forwarded-For spoofing Closed
Origin-bypass escalation Closed
Dev-login endpoint bypass Closed

Your AI should protect you,
not the other way around.

As AI gets more capable, security architecture matters exponentially more. Nebo was built for that future.