Attenuating Permissions from Humans to Autonomous Agents via Cryptographic Capabilities

Core engineering and protocols for Verifiable Delegation Chains.

The Principal-Agent Problem in Autonomous AI

When a human user instructs an autonomous executive assistant to 'book a flight and hotel for under $2,000,' the user delegates real-world financial authority. If the agent is tricked by a prompt injection attack on a hotel website, it might be instructed to transfer the entire corporate treasury. Capability delegation must be cryptographically bounded.

Attenuated Delegation Tokens with Post-Quantum Proofs

A2AShield utilizes modern capability tokens (similar to Macaroons and Biscuits) signed with ML-DSA. The human signs the initial root token, embedding strict constraints: (Max_Spend = $2,000, Purpose = Travel, Expires = 24h). The primary agent can delegate sub-tasks to specialized travel search agents, adding further restrictions, but can never expand the original permission envelope.

Explore Technical Architecture & Portfolio