Agent Execution Isolation
Keeps the project on the host and routes only agent operations through Docker — minimising blast radius and protecting workspace files.
SECLAW SECURED AGENT · V0.1.0
SeClaw —
Secured Personal AI Assistant
Built for safe execution first: isolated actions, strict tool-call validation, prompt-injection defense, risky-output sanitization, and one-command rollback when anything goes wrong.
Launch
Install and run SeClaw in minutes — locally or with Docker isolation.
Launch in minutes
Install dependencies, bootstrap config, and run gateway.
npm install -g seclaw-agent
seclaw onboard
seclaw gatewayRun in Docker
Enable agent execution isolation by running operations inside a Docker sandbox. Download Docker Desktop, keep it running, then set security.dockerSandbox.enabled to true in your config.
seclaw onboard
# set security.dockerSandbox.enabled = true
seclaw gatewaySecurity Layers
10 security mechanisms protect your agent procedure.
Snapshot & Rollback
CoW mechanics snapshot and restore mounted files in seconds, letting the agent work freely with a guaranteed recovery path.
Prompt Injection Defense
CFI/IFI at system level constrains valid action space; a guard model sanitises suspicious tool outputs before they re-enter the reasoning loop.
Skill Audit
Scans skills for dangerous patterns — prompt injection, data exfiltration, and destructive shell commands.
Memory Audit
Scans memory files for stored prompt-injection payloads, leaked credentials, and PII exposure risks at rest.
Execution Audit
Records full task traces and generates a post-execution risk report flagging potentially dangerous actions after each run.
Privacy Protection
Monitors identity info, API keys, SSH keys, and credentials in real time — flagging exposures before they leak through tool outputs or channels.
Risky Operation Protection
Detects dangerous commands such as rm -rf or sudo and requires explicit user confirmation before execution.
Secure Communication Isolation
Maintains separate context windows per channel, blocking cross-channel prompt injection and inter-source manipulation.
Network Security Controls
Enforces HTTPS, applies request timeouts, and supports configurable network modes to prevent uncontrolled external access.
Risk Coverage
SeClaw protects against 8 of the Top 10 OWASP 2026 Agent Risks. The remaining two risks, Insecure Inter-Agent Communication and Human-Agent Trust Exploitation, do not apply to the current Claw-based agent workflow.
01Agent Goal Hijack
02Tool Misuse
03Identity & Privilege
04Supply Chain
05Code Execution
06Memory Poisoning
08Cascading Failures
10Rogue Agents