The Problem
AI agents that can execute code need guardrails: what commands can they run, what files can they access, how much CPU/memory can they consume, and can they phone home?The Solution
What Each Layer Does
| Layer | Protects Against |
|---|---|
Glassbox validate_command | Dangerous shell patterns (rm -rf, sudo) |
Ironclad validate_command_deep | Piped attacks, reverse shells, crypto mining |
Ironclad sandboxed_command | Network access, filesystem escape, privilege escalation |
Ironclad spawn_watchdog | CPU/memory abuse, infinite loops |
Glassbox validate_output | System prompt leaks in output |

