Security Architecture

A deep dive into how we protect your AI agent. For a quick overview, see our Security page.

Infrastructure isolation

Five layers of network and hardware isolation that make your agent invisible and inaccessible.

Outbound-Only Encrypted Tunnels

Every agent connects to the outside world exclusively through outbound-only encrypted tunnels with TLS 1.3. Zero listening ports on your server. Invisible to port scanners. No one on the internet can initiate a connection to your agent.

DM Pairing (Contact Verification)

Before your agent responds to a new contact on any messaging platform, it requires your explicit approval. This blocks social engineering attacks and unauthorized access.

LUKS2 Encrypted Storage

Storage volumes are encrypted at rest using LUKS2 with AES-256-XTS. Encryption keys are managed separately from the storage volumes. Physical access to hardware cannot expose your data.

No Incoming Traffic

The agent gateway binds to localhost only. All inbound connections from the public internet are blocked. Your agent can make outbound requests, but nothing external can reach it.

Isolated Virtual Private Server

Every agent runs on its own dedicated VPS. No multi-tenancy at the infrastructure level. CPU, memory, storage, and network are completely isolated from every other customer.

Agent EDR: Runtime security

New

Six detection layers that run outside the agent, at the kernel level and control plane. Even if your agent is compromised, these layers survive.

Tool Execution Policies

Every tool invocation is evaluated against our security policy engine before execution. Dangerous operations (reverse shells, crypto miners, credential theft, privilege escalation) are blocked or flagged in real-time. Every tool call is logged.

File Integrity Monitoring

Critical system files, agent configurations, skill files, and binaries are monitored continuously on the host OS. Any unauthorized modification, creation, or deletion is detected and alerted in real-time. Runs outside the agent.

eBPF Runtime Monitoring

Kernel-level eBPF probes monitor every process spawn, file access, and system call on the host. Catches crypto miners, reverse shells, unauthorized binaries, and data exfiltration attempts. Completely invisible to and unreachable by the agent. Available on Pro and Enterprise plans.

Secrets Isolation

Credentials are fetched on-demand through authenticated channels and cached in memory only. API keys are never written to disk on the agent VPS, eliminating the risk of credential theft through environment variable or filesystem access.

Network Threat Intelligence

Every outbound connection is checked against threat intelligence feeds updated daily. Known malicious IPs (botnet command-and-control, malware infrastructure) are blocked at the firewall level before the connection is established.

Behavioral Anomaly Detection

Our behavioral engine establishes a baseline of normal activity for each agent: tool usage patterns, network destinations, resource consumption. Sudden deviations trigger investigation. Coming soon.

Detection architecture

Security layers run at three levels. Only the tool policy engine runs inside the agent. Everything else runs on the host or control plane, outside the agent's reach.

Control Plane

Alert DatabaseSecurity PoliciesBehavioral Engine

Host Operating System

eBPF Runtime MonitoringFile Integrity MonitoringThreat Intel Firewall

Agent Container

Tool Policy EngineOpenClaw + Skills

Network isolation model

Your agent's network architecture eliminates entire classes of attacks.

Public Internet

Blocked

Edge Network

TLS termination

Encrypted Tunnel

Outbound-only

Agent VPS

Localhost-only

Zero listening ports. Invisible to Shodan, Censys, and all port scanners
All connections initiated outbound by the agent, never inbound
Defeats port scanning, brute force, and direct exploitation attacks

Encryption

Three layers of encryption protect your data at every stage.

At rest

AES-256-XTS encrypted storage volumes (LUKS2)

In transit

TLS 1.3 encrypted tunnels for all communication

Field-level

AES-256-GCM for sensitive database fields (tokens, secrets)

Token security

Credentials are handled with defense-in-depth at every step.

Timing-safe comparisons prevent attackers from guessing tokens through response timing differences
Credentials stored in an encrypted vault on the control plane, never written to disk on the agent VPS
Credential rotation API allows rotating all tenant secrets without downtime
All webhook payloads verified with HMAC signatures using fail-closed validation

Audit logging

Every action is recorded with full context for traceability.

Every tool invocation logged with timestamp, input, and output
Authentication events tracked (login, token refresh, session expiry)
Configuration changes recorded with before/after values
Tamper-evident log storage: modifications are detectable
Configurable retention policies per compliance requirements

Ready to hire your first AI employee?

Production-ready in 5 minutes. All 7 security layers included on every plan.