Adversarial AI Agents for Web3 Security
summary
Argues for RL agents that model attackers using an MDP — states → actions → rewards — and defines invariants as win conditions; a tested system called Red-Team Copilot trains on re-entrancy and then autonomously discovers a new timestamp bug, showing zero‑day find capability. Notes LLM agents can be prompt‑injected (one incident led to a wallet transfer of ~$100k after ~40 successful manipulations) and presents QuillGuard as guardrails for agent behavior. The approach is presented as a continuous adversarial layer alongside audits, with an implemented prototype and a research paper.
AI-written summary · about
chapters (15)
- 0:00Start
- 0:02Opening & background (QuillAudits, 1,500+ protocol engagements)
- 1:37Constant in Web3: weekly hacks/scams despite growth
- 2:13Trust deficits: founders/devs/third parties; broader attack surface (bridges, AMMs, lending)
- 3:15Why current tooling falls short: static, symbolic, fuzzing limits
- 5:20LLM agents are attackable: prompt injection, memory wipe, nondeterminism
- 6:54Real incident: agent manipulated to transfer funds (~$100k)
- 8:04Monitoring/AI “auditors” trained on past bugs struggle with zero-days
- 9:41Enter adversarial RL: self-learning attacker agents
- 10:42MDP framing: state, action, reward; define invariants as win conditions
- 11:48Applying to contracts & agents; reward signals and learning loops
- 13:05Red-Team Copilot demo concept: access-control & multi-tx paths
- 14:14Result: RL trained on re-entrancy later found a timestamp vulnerability unprompted
- 14:52Extending to AI agents: QuillGuard for guardrails
- 15:25Call to action & research paper; future of self-learning adversaries
related talks
19mKeynote: Towards Building Safe and Secure Agentic AI
27mKeynote: Towards Building Safe and Secure Agentic AI
26mHow Bad Actors AbUse LLMs The Mechanisms That Enable Misuse and What We Can Do About
33mSecuring Agentic AI With PyTorch: Threat Modeling & LLM Red Teaming in Practice
01h 01mIntroduction to AI Security
31mAgentic AI Threats: How Multi-Agent Systems Get Exploited
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
42mSecuring AI Agents...