An Empirical Case Study of Prompt Engineering Strategies for Bypassing Multi-Layer Agent Security Defenses

arXiv:2610.14116 · cs.CR · Submitted 2026-10-09 · Read on arXiv

cs.CR

Submitted: 2026-10-09

Updated: 2026-10-09

Comments: 5 pages, 6 tables. Accepted at The 2nd Workshop on AIGC Safety from Dual Perspectives of Model and Content (WS06), IJCAI-ECAI 2026. Case study from Track 1 of the OpenClaw Security Attack and Defense Challenge

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Prompt Engineering — 1× in this paper · explained in 5 episodes

Sources

• SeClaw: Spec-Driven Security Task Synthesis for Evaluating Autonomous Agents
• Prompt Injection attack against LLM-integrated Applications
• Universal and Transferable Adversarial Attacks on Aligned Language Models

Related papers

• SoK: AI-Augmented Binary Reversing
• Relaxed Sender Anonymity for CBDC Interbank Settlement: A Zero-Knowledge Approach on Permissioned EVM
• Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
• Estimating the Social Cost of Corporate Data Breaches
• Efficient Fuzzy PSI under One-Sided Assumptions
• Sealing the Audit-Runtime Gap for LLM Skills