ARC Loop is not a sandbox.
It checks cooperative agent behavior through contracts, hooks, and validation. It does not isolate untrusted code.
Do not use it as a replacement for containers, VMs, OS sandboxing, CI permissions, or code review.
ARC helps with:
- agents editing files outside the agreed scope
- agents running blocked shell commands through normal tool paths
- agents changing Done criteria after launch
- agents claiming completion without verified evidence
- repeated failed attempts that should pause
ARC does not protect against:
- malicious code executed outside the host hook system
- compromised dependencies
- shell commands that hide harmful behavior inside allowed wrappers
- weak verifier commands
- a human accepting a bad amendment
Open a GitHub issue with:
- the ARC version
- the host agent, Codex or Claude
- the relevant
arc.toml - the hook input if the bug is hook-related
- the command you ran
- expected and actual behavior
Do not include secrets from .env, logs, prompts, or agent transcripts.