Opaque

As demonstrated by Patlan et al. (Cornell / UIUC, 2025) in “Real AI Agents with Fake Memories: Fatal Context Manipulation Attacks on Web3 Agents” , Dominant agent architecture today is fundamentally unsafe, and the problem is not the smart contract, but agent context. Modern Web3 agents are vulnerable to context manipulation attacks, including:

1. prompt injection

2. memory poisoning

3. persistent context corruption
   
   
   Our Solution

   Opaque is a policy-enforced execution layer for Web3 agents where:

   • Agents propose actions

   • Policies are evaluated outside the agent’s context

   • Authorization is proven cryptographically

   • Execution happens only after on-chain verification
     

The agent never sees policy logic, thresholds, constraints, or decision rationale. The Cornell research showed that:

• Attacks succeed because policies live in agent context.

• Memory persistence makes compromise durable

• Prompt defenses fail
  
  

Opaque eliminates the attack surface by removing policies from agent context, enforcing them via cryptographic proof, and anchoring verification on-chain before execution. Mantle Network provides low-cost, fast verification and execution layer for the agents intended actions