Adversarial Poetry Can Jailbreak LLMs — Even in a Single Prompt

The paper shows that rewriting harmful or restricted requests as poetry (“adversarial poetry”) can reliably bypass safety mechanisms in large language models (LLMs). Across 25 state-of-the-art models, hand-crafted poetic prompts achieved an average “attack success rate” of 62%, with some models exceeding 90% success.

Even when standard harmful prompts from a broad safety benchmark were converted automatically into verse, the poetic versions produced up to 18× higher success rates than their prose originals — showing this vulnerability isn’t limited to a few handcrafted examples. 

The authors argue that poetic style — its metaphors, rhythm, and unconventional structure — alone suffices to evade guardrails across many risk domains (cyber-offense, harmful manipulation, dangerous instructions, etc.), exposing a systemic weakness in current alignment and safety-evaluation practices. 

https://arxiv.org/html/2511.15304v1

Comments

Post a Comment

Popular posts from this blog

Prompt Engineering Demands Rigorous Evaluation

In the blog post “Prompt Engineering Requires Evaluation” on the Shostack + Associates website, the author argues that treating prompts for large language models (LLMs) merely as creative artefacts is insufficient. Engineering prompts properly demands structured evaluation frameworks — what the AI community calls “evals” — to test which prompt versions work better, with which models, and under which conditions. The post highlights that simply assuming a prompt is “good enough” creates risks when LLMs are integrated into production systems (e.g., for threat modeling). It advocates for measuring prompt performance, variation effects, and tool-chain dependencies (model, context, ancillary materials).  Ultimately the message is: prompt engineering should borrow disciplined practices from software engineering (versioning, testing, benchmarking) rather than relying on informal experimentation.  https://shostack.org/blog/prompt-enignieering-requires-evaluation/
Read more

Secure Vibe Coding Guide: Best Practices for Writing Secure Code

Ken Huang's "Secure Vibe Coding Guide" emphasizes the importance of integrating security into the software development lifecycle. The guide provides best practices for writing secure code, including input validation, proper authentication mechanisms, and secure data storage techniques. It also highlights the necessity of regular code reviews and staying updated with the latest security vulnerabilities and patches. By following these guidelines, developers can create applications that are resilient against common security threats and contribute to a safer digital environment.  https://kenhuangus.substack.com/p/secure-vibe-coding-guide
Read more

KEVIntel: Real-Time Intelligence on Exploited Vulnerabilities

KEVIntel is a dynamic platform providing up-to-date information on known exploited vulnerabilities (KEVs). It aggregates data from over 50 public sources, including CISA, and enriches each entry with metadata such as EPSS scores, online mentions, scanner inclusion, and exploitation status. The platform aims to serve as an early warning system, offering insights even before official publications. KEVIntel supports various formats like JSON, CSV, and RSS, facilitating integration into security operations. These entries include CVSS scores, exploit status, and links to proof-of-concept code, aiding organizations in prioritizing remediation efforts.   https://kevintel.com/
Read more