Deepening ambiguity: A*-inspired multi-agent commonsense obfuscation attacks against LLM prompts

Researchers propose an A*-inspired multi-agent commonsense obfuscation attack framework for generating semantically aligned but obfuscated LLM prompts. Its hierarchical rewriting strategy is guided by a dynamic semantic dispersion coefficient γ, follows a reverse simulated annealing schedule with conservative edits early and aggressive obfuscation later, and uses Agentic Mechanism Labeling to refine the adversarial mechanisms. Across multiple LLMs, the method achieves a higher attack success rate than exhaustive exploration while requiring fewer attempts.