How ArmorClaw enforces security
mainArmorClaw operates in two main phases to protect AI agents:
1. Intent Planning
When a message is sent to an OpenClaw agent, ArmorClaw:
- Intercepts the LLM input via the
llm_inputhook. - Parses available tools from the system prompt.
- Generates an explicit plan of allowed tool actions via a separate LLM call.
- Sends the plan to the ArmorClaw backend to receive a cryptographically signed intent token.
2. Tool Execution Enforcement
Before any tool is executed, ArmorClaw:
- Verifies the tool is in the approved plan.
- Validates the intent token is not expired.
- Applies local policy rules.
- Optionally verifies CSRG cryptographic proofs.
- Blocks execution if any check fails.