https://docs.openparallax.dev

I kept reading about security issues with open source AI agents. Deleting important data, exfiltrating confidential information. Another thing was the technical difficulty of setting one up. Non-technical friends of mine kept reaching out because they were unable to find their way around installing it.

I built an open-source AI agent for personal use where the LLM reasoning system literally cannot execute anything. It runs in a sandboxed process with no filesystem, no network, no shell. It proposes actions over gRPC to a separate engine process that validates every action through four tiers. YAML policy rules, a heuristic engine + DeBERTa classifier, an independent (reversed prompted) LLM evaluator, and human in the loop. Actions that fail validation are blocked. The system also tags data sensitivity (IFC) and snapshots state before destructive actions for rollback. Multiple heterogeneous levels of validation means a wider coverage of attack surfaces. The reasoning system is treated as fully compromised by default, so the security boundary holds regardless of what the LLM does.

I wrote it in Go, it's one single static binary. Tested it against 280 adversarial cases with 98.9% block rate and zero false positives under default config. The agent is multiplatform, installed with one command, setup and ready to run in under 3 minutes (with API keys in hand). No external dependencies needed (no nodejs, python or docker)

Repo: https://github.com/openparallax/openparallax

Paper (draft, arXiv submission pending): https://github.com/openparallax/openparallax/releases/downlo...

Happy to answer questions about the architecture, threat model, or evaluation.