Apppublic

Ayush-Kumar0207/panopticon-protocol-v3

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes

App README

👁️ The Panopticon Protocol v3

Counter-Espionage RL Environment — "Among Us… for AIs"

An OpenEnv-compliant reinforcement learning environment where an AI security agent (ARGUS) defends a corporate network against an adaptive adversary (HYDRA) that infiltrates multi-generational sleeper agents.

What Makes This Unique

7 Stacking Espionage Mechanics that create emergent strategic depth:

Mechanic	Description
Canary Traps	Plant unique tracking data to trace leaks
5-Gen Sleepers	From amateur leakers to Manchurian candidates
False Flags	Gen-3+ sleepers frame innocents
Dead-Man's Switches	Gen-4 explode on careless termination
Double Agent Turning	Convert enemies into YOUR assets
Disinformation	Feed false intel back to the adversary
HYDRA Adaptive Memory	The enemy evolves its counter-tactics

6-Phase Narrative Arc

Orientation → First Contact → Escalation → Deep Cover → Crisis → Counterstrike

Why This Matters

This environment trains the exact capabilities needed for Scalable AI Oversight:

—Theory-of-Mind: Modeling hidden states and intentions of other agents
—Deception Detection: Identifying false signals in noisy, adversarial data
—Strategic Planning: Multi-step reasoning under partial observability
—Adaptive Response: Adjusting tactics as the adversary evolves

API Endpoints

—GET /health — Health check
—POST /reset — Start a new episode
—POST /step — Take an action
—GET /tasks — List available difficulty levels
—POST /grade/{level} — Grade an episode

Quick Start

bash

pip install -r requirements.txt
uvicorn _server:app --host 0.0.0.0 --port 8000

License

Apache-2.0