Towards Interactive Research Agents for Internet Incident Investigation

We explore interactive simulacra of human researchers for Internet incident investigation. Building on Auto-GPT and GPT-4, our agent autonomously retrieves information, maintains knowledge memory, and self-learns to investigate complex Internet incidents. On challenging scenarios such as hypothetical solar-storm impacts on networks, the agent reaches 87.5% consistency with human expert insights, demonstrating that LLM-based agents can meaningfully automate parts of the investigation process.

In Proceedings of the 22nd ACM Workshop on Hot Topics in Networks (HotNets ‘23)