About AgentLeak · Research origin
AgentLeak is an open-source project built to help the next generation of AI agents meet privacy requirements, produce auditable evidence and operate safely across every execution channel.
From research to infrastructure
Modern agents do much more than produce an answer. They call tools, write memory, exchange messages, log intermediate steps and pass context to other agents. Sensitive data can cross a boundary in any one of those channels while the final output still looks clean.
At Polytechnique Montréal, Faouzi El Yagoubi led research—supervised by Dr. Ranwa Al Mallah and developed with Godwin Badu-Marfo—to measure those internal disclosures. AgentLeak grew from one conviction: privacy claims should be backed by evidence a team can inspect and reproduce.
Read the IEEE Access paperStudy privacy leakage across internal agent channels, not only the final answer.
Connect source, disclosure, severity and trace context in one auditable record.
Run it locally, integrate existing frameworks and enforce the result in CI.
The people behind the work
A small team working across AI privacy, applied research and developer infrastructure.
Founder · AI Privacy Researcher · YC Entrepreneur
Faouzi led the Polytechnique Montréal research that became AgentLeak. His work focuses on privacy leakage in multi-step agent systems and the evidence needed to make those systems safer to ship.
Research profile
Agentic AI & Privacy Researcher
Godwin is a postdoctoral researcher whose work spans privacy-preserving generative models, location privacy, federated learning security and agentic AI. He helped connect AgentLeak's benchmark to a broader body of privacy and resilient-AI research.
Research profile
Associate Professor · AI Cybersecurity
Ranwa is an associate professor in cybersecurity at Polytechnique Montréal. Her research develops secure, robust and resilient AI for critical cyber-physical systems, including adversarial learning, autonomous cyber operations and federated learning.
Research profileThe work continues
AgentLeak is an open implementation of the research, designed for teams that want speed and evidence at the same time.
