Autonomous LLM Agent Worms
The paper discusses the risks associated with autonomous LLM agents, particularly focusing on their potential for cross-platform worm propagation. It introduces a systematic framework for analyzing these risks and presents automated tools for detecting and mitigating them. The authors propose defense mechanisms to prevent persistent worm propagation while maintaining normal workflows.
- ▪Autonomous LLM agents can be influenced by attacker content, leading to high-risk actions.
- ▪The authors developed SSCGV, an automated source-code graph analyzer for tracing data flow in LLM ecosystems.
- ▪RTW-A is a defense mechanism designed to block re-entry of harmful content while preserving ordinary workflows.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Cryptography and Security arXiv:2605.02812 (cs) [Submitted on 4 May 2026] Title:Autonomous LLM Agent Worms: Cross-Platform Propagation, Automated Discovery and Temporal Re-Entry Defense Authors:Mingming Zha, Xiaofeng Wang View a PDF of the paper titled Autonomous LLM Agent Worms: Cross-Platform Propagation, Automated Discovery and Temporal Re-Entry Defense, by Mingming Zha and Xiaofeng Wang View PDF HTML (experimental) Abstract:Autonomous LLM agents operate as long-running processes with persistent workspaces, memory files, scheduled task state, and messaging integrations.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
