PIC Standard: Bridge the Causal Gap in Agentic AI
Imagine your AI agent, smart, autonomous, handling real tasks like booking flights or processing payments, suddenly goes rogue because of a sneaky untrusted input, like a manipulated prompt. That's the "causal gap": flawless reasoning leading to unintended disasters.
In a world where AI agents are exploding in use (think LangGraph or CrewAI), we need simple, reliable safety nets.
PIC Standard (Provenance & Intent Contracts) is an open-source protocol that makes agentic AI safer without the hassle.
Before any tool runs (ie. an API call or data sync), the agent must submit a JSON "proposal" that answers three key questions:
Provenance: Where did the data come from? (Trusted database? Semi-trusted API? Untrusted user input?)
Intent: Why this action? (Clear rationale to avoid mishaps.)
Impact: What's the risk? (A simple taxonomy: low like "read," medium like "write," or high like "money" for finances or "privacy" for personal data.)
If the proposal doesn't check out, boom, blocked automatically. No more accidents.
What sets PIC apart? It is inspired by advanced research like Google DeepMind's CaMeL (for secure dialogues) and RTBAS (for robotic safety), but built for everyday devs: Lightweight JSON schemas, a Python SDK you can pip-install, and drop-in integrations (ie. a single node for LangGraph workflows).
No custom interpreters or heavy dependencies, just practical guardrails.
New in v0.3.0: Verifiable Evidence! Now, trust is not just declared, it is proven.
Agents reference real artifacts (like files), and PIC auto-verifies hashes to catch tampering. For high-stakes actions, it fails closed: No verified proof? No go.
It ships with LangGraph and Anthropic's MCP (Model Context Protocol) integrations.
Whether you are in FinTech (securing payments), SaaS (protecting user data), or robotics (ensuring safe commands), PIC bridges those gaps elegantly.
It is Apache-2.0 licensed, extensible (add your own risk classes), and community-driven. Star, fork, or contribute!
Roadmap ahead: Signatures for endorsements (v0.4) and API attestations for external checks (v0.5).
Join the movement to make AI agents trustworthy.
Check it out: https://github.com/madeinplutofabio/pic-standard
Questions? Let's chat!
