Can We Verify Every Action of an AI Agent? |

Verifiable AI agent playing Doom with cryptographic proofs for every action

🎮 Can We Verify Every Action of an AI Agent?

Excited to announce our innovative solution: The first verifiable AI agent playing Doom, achieving near real-time performance and scalability for sequential decision making! In the video, every action comes with a cryptographic proof and the entire action space and agent operate in a verifiable circuit. The agent makes autonomous decisions while providing proof of inference, with model weights potentially remaining hidden from the tracking system yet verifiable.

💡 With the rise of Agentic AI, innovation, strategic improvement and driving efficiency in business transformation are inevitable. However, AI agents must be trustworthy to participate in autonomous decision making. AI agents inherently introduce new risks requiring security, verifiability and governance structure across three crucial layers:

Human-agent interaction: Implementing AI agent traceability and interpretation mechanisms for human verification.

Agent integrity: AI agent accountability and data accessibility to avoid alignment faking and hallucination.

Multi-agent systems: Design of validator agents in multi-agent systems or sharing protocols between workflow of agents needed (equivalent to HTTPS protocol).

🎯 AI agent structure must be private, yet interpretable, fair and transparent in behavior. This innovative solution, though in early stages, can immediately improve human feedback loops, discover new training methods and eliminate biases!

Some applications include:

Ensure LLM outputs are unbiased and accountable through fair evaluation methods
Verify AI medical diagnoses to confirm all vital signs (like blood pressure) are properly considered
Enable antifraud systems with private and secure authentication protocols
Secure knowledge transfer between AI agents while preventing collusion and adversarial attacks
Verify model inference integrity in third-party computation processes
Provide verifiable proof of model predictions to avoid liability in legal and insurance requirements
Monitor and verify 24/7 AI assistant performance and task completion
Establish verifiable benchmarks for comparing different AI models

💫 Inference should be verifiable! We need private, accountable online interactions equipped with zero-knowledge proofs for seamless, traceable decision-making. A cherry on top: zero-knowledge cryptography is post-quantum ready. 🍒

🤝 This Inference Labs project comes to life through the efforts of our team, visionary idea from Ron Chan and Colin Gagich, led by Spencer Graham, PoC by Shirin Shahabi with collaboration of Eric Lesiuta, Shawn Knapczyk, Oleh Bondarenko, Tristan Freiberg, Jonathan Gold, Noel Prangley and Julia Théberge in inference lab team.

Let’s shape the future of AI governance together! Interested in collaboration or have technical questions? Let’s connect!

Originally posted on LinkedIn