Projects

Secure Infrastructure for Agentic Systems

Designing the underlying architectures, protocols, and trust boundaries that enable users to control and manage their agents.

• SAGA
• MAGIQ
• SAGA-BFT

Web Agent Security

Studying how web agents can be misled, hijacked, or otherwise driven toward unintended behavior via indirect prompt injection — and how to constrain them.

• MUZZLE

Defenses for Agentic Systems

Developing mitigation mechanisms against agentic attacks, such as indirect prompt injection, using security principles.

• ACE

Multi-Agent Systems

Designing architectures that enable multiple agents to securely interact and collaborate in solving complex tasks.

• APWA

RL Security

Attacks and defenses against reinforcement-learning agents, including training-time backdoor and poisoning threats.

• SleeperNets
• DAZE
• Q-Incept

Automated Cybersecurity Defenses

Using LLM agents to scale up security analyst capacity for incident response, threat investigation, and mitigation planning.

• LLMs for Security Incident Analysis