How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense
Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in…
AI Summary & Key Takeaways
- Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper
- We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in…
How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense
Related stories
Anthropic takes on Grok Bot, Meta's Muse, and OpenAI's Dots with its always-on agent "Claude Duet"
Anthropic has launched Claude Duet, a proactive AI agent that handles tasks on its own, even when the user is offline. The app is built on Claude Opus 5.5 and connects to services like Gmail, Slack, and Stripe through more than 6,000 connectors. It asks for approval before…
Ensemble forecasting with IBM Granite time-series models
A Blog post by IBM Research on Hugging Face
In ChatGPT you can hide conversations using a PIN code or Face ID
🔒 In ChatGPT you can hide conversations under a PIN code or Face ID OpenAI has begun to gradually launch the Locked Chats feature, which allows you to protect individual conversations in ChatGPT from strangers. Users can hide personal or confidential chats from regular history and...
RUЕвропейские пионеры используют модели NVIDIA Earth-2 для прогноза погоды
Модели NVIDIA Earth-2 помогают прогнозировать погоду и климат в Европе.