This research reveals how situation-driven attacks using everyday contexts can create deceptive prompts that evade LLM safety mechanisms.

• Uses realistic movie script scenarios to craft human-readable adversarial prompts
• Creates prompts that appear harmless to humans but trigger harmful LLM responses
• Demonstrates higher success rates than nonsensical attacks, which can be easily detected
• Highlights critical security gaps in current LLM defense strategies

For security teams, this work exposes a concerning vulnerability: attacks that blend into normal conversation are significantly harder to detect and filter, requiring new approaches to LLM safety.

Human-Readable Adversarial Prompts: An Investigation into LLM Vulnerabilities Using Situational Context