Research Intern

Conducted research on indirect prompt injection vulnerabilities in LLM-based agentic systems.

• Developed a gym-like environment for scalable training-data sampling and generation, supporting online and offline post-training of assistant models and standalone safety verification models.
• Trained LLMs from multiple model families using supervised fine-tuning and reinforcement learning methods, including DPO and GRPO.
• Achieved state-of-the-art performance on established benchmarks, including AgentDojo and AgentDyn.