Home / Posts / Agent Trust Protocol

Agent Trust Protocol (ATP): Solving the "Who Am I?" Problem for Autonomous AI

Agent Trust Protocol

As we enter the summer of 2026, the internet is no longer just a place for humans and bots; it is a dense mesh of autonomous agents executing financial transactions, modifying codebases, and managing infrastructure. On May 12, **Lyrie.ai** (formerly OTT Cybersecurity) unveiled the Agent Trust Protocol (ATP), the world's first open cryptographic standard for verifying agent identity and authority.

Why Traditional Auth Fails for Agents

OIDC and OAuth were designed for humans interacting with apps. Agents, however, operate recursively. An agent might spin up three sub-agents, which then interact with external APIs. Traditional bearer tokens are easily stolen and lack the contextual provenance needed to prove that a specific action was authorized by the parent system.

How ATP Works: The Cryptographic Proof

ATP introduces the concept of a "Trust Envelope." Every request sent by an ATP-compliant agent contains a non-interactive Zero-Knowledge Proof (zk-SNARK) that verifies:

  • Origin: The parent LLM runtime that spawned the agent.
  • Policy: A hash of the specific `system_prompt` and constraints the agent is operating under.
  • Ephemeral Key: A short-lived cryptographic key bound to the specific task session.

The "Agent Passport"

"ATP gives every agent a passport. When an agent hits your API, you don't just see a token; you see a verifiable proof of its lineage and its permitted scope of action." — Lyrie.ai Engineering

Broad Industry Adoption

Major infrastructure providers including **Cloudflare**, **Akamai**, and **Vercel** have announced day-one support for ATP at the edge. By verifying ATP proofs at the CDN layer, providers can block "rogue" or "shadow" agents before they ever reach the application logic, significantly reducing the attack surface for agentic exploits.

The Future of Verifiable Compute

The release of ATP marks a critical step toward **Verifiable Compute**. In a world where AI agents handle trillions of dollars in economic value, the ability to prove *what* an agent is and *why* it is performing an action is no longer optional—it is a foundational requirement for the agentic economy.