← All services
what we do
Security & Guardrails
The controls that make an autonomous agent safe to run against real systems and real money.
The problem
An agent with tools can take actions that cost money, change records, or message customers. Without hard controls, one bad plan or one prompt-injected input is a production incident.
Security can't be a prompt asking the model to behave. It has to be enforced outside the model, in the runtime and the tool layer.
What we build
- Authentication, and per-agent, per-tool authorization.
- Sandboxed execution — no ambient network or filesystem access.
- Spend limits and rate limits enforced at the gateway.
- Human-in-the-loop approval gates for irreversible actions.
- Prompt-injection defenses at the tool boundary and on untrusted inputs.
- Immutable audit logs of every action, input, and decision.
How it works
- Classify every tool by blast radius — read, draft, or act.
- Put approval gates in front of the “act” tier.
- Sandbox the runtime and scope its credentials.
- Turn on audit logging and spend caps before the first live run.
What you get
- Authorization model and sandbox configuration.
- Approval-gate workflow for high-risk actions.
- Audit log pipeline and retention policy.
Run it your way
Every build ships one of two ways. Most engagements start managed and move in-house once the team is ready — some stay managed for good.
how we build
- Controls enforced in the runtime and MCP layer, not in prompts.
- OpenClaw's gateway treated as a security boundary; it is never exposed directly — reached over an SSH tunnel or Tailscale.
- Spend caps, quotas, and kill switches at the gateway; audit logs shipped to your SIEM or log store.