← All programs
AAS-301AdvancedAI38h · 5 weeks (self-paced)

AI Agent Security for Web3

The hottest attack surface in crypto today. Autonomous AI agents are managing wallets, executing swaps, and interacting with protocols — and the security implications are severe. This program covers both offensive (how agents get exploited) and defensive (how to build them safely): prompt injection leading to fund loss, agent key management, guardrails for on-chain actions, and monitoring autonomous systems.

01 · Outcomes

What you will be able to do.

  • 01Threat-model an AI agent that holds signing keys and interacts with DeFi
  • 02Demonstrate prompt-injection attacks that lead to unauthorised on-chain actions
  • 03Design defence-in-depth guardrails: allowlists, simulation, spend limits, kill switches
  • 04Implement secure key management for autonomous agents (enclaves, MPC, session keys)
  • 05Build a monitoring and alerting pipeline for agent-initiated transactions
02 · Syllabus

What you will cover.

Week 1

AI Agent Architectures in Web3

  • Anatomy of an on-chain AI agent: planner, tool-caller, signer
  • Survey of deployed agent frameworks (Eliza, GAME, custom)
  • Threat modelling an agent that can move funds
Week 2

Prompt Injection & Agent Exploitation

  • Direct and indirect prompt injection in agent contexts
  • Exploiting tool-use: making an agent call the wrong contract
  • Data exfiltration through agent observation channels
Week 3

Guardrails & Policy Enforcement

  • Transaction simulation before signing (Tenderly, local fork)
  • Allowlists, spend limits, and rate limiting at the signing layer
  • Kill switches and human-in-the-loop escalation
Week 4

Key Management for Autonomous Systems

  • Session keys, scoped permissions, and ERC-7715 / ERC-7710
  • TEE and enclave-based signing for agents
  • MPC-based agent wallets and threshold approval
Week 5

Monitoring, Incident Response & Capstone

  • Real-time monitoring of agent-initiated transactions
  • Alert pipelines and automatic pause triggers
  • Capstone: red-team an AI agent and write the defence report
Tools
  • • Tenderly
  • • Safe{Wallet}
  • • OpenZeppelin Defender
  • • LangChain
  • • Anthropic Claude
Who it's for
  • Builders deploying AI agents that interact with smart contracts
  • Security teams reviewing agent-based products
  • Protocol teams integrating AI-agent APIs into their platforms
Prerequisites
  • Completion of CSF or equivalent wallet and key-management knowledge
  • Understanding of how LLMs and tool-use / function-calling work
  • Basic Solidity or smart-contract interaction experience
Ready?

Enrol in AI Agent Security for Web3