Yang Xianglin

Research Fellow @NUS | Ph.D. @NUS | working on AI Safety and XAI

[Yang Xianglin @xianglin_yang]

What we found 👇

🧠 Agent memory can turn one-time prompt injections into persistent compromises
🔁 Attacks can survive across sessions in sliding-window and RAG-based agents
🛠️ Compromised agents may later misuse tools or leak sensitive data

[Yang Xianglin @xianglin_yang]

Can self-improving AI agents become a security risk? 🧠

In our new paper, we study how LLM agents that learn from experience can also absorb malicious instructions from untrusted content, creating persistent safety risks across sessions. 🧟‍♂️🤖

🔗 Paper: https://t.co/0mntIrGMpm