Show HN: Semantic Overlays – an NX bit for LLM prompt injection (live demo)

Joshua Penman released Semantic Overlays, a method using small trained adapters on a frozen model to mitigate LLM prompt injection, claiming it brings the injectable Qwen-3.5-9B to state-of-the-art scores on all prompt injection benchmarks he could find. The paper, code, and adapters are available on arXiv, GitHub, and Hugging Face, with a live demo inviting users to attempt to break the system.

I've built a new method for steering LLMs called Semantic Overlays, small trained adapters on a frozen model which change how its perceives a piece of its context. The most readily applicable usage is to mitigate prompt injection, and it lets us take a very-injectable Qwen-3.5-9B to SOTA scores on all the prompt injection benchmarks I could find. They are only blackbox attacks, but I did NOT train on anything like them — whitebox attacks are out of scope for this paper I'm excited for you to play with the tech — see if YOU can break it let me know if you can Paper at https://arxiv.org/abs/2608.23873 https://arxiv.org/abs/2608.23873 if you want to read more about it, code at http://github.com/JoshuaSP/semantic-overlays http://github.com/JoshuaSP/semantic-overlays , adapters at http://huggingface.co/joshuapenman/semantic-overlays-adapter... http://huggingface.co/joshuapenman/semantic-overlays-adapters Also https://x.com/joshua s penman/status/2094823990472884389 https://x.com/joshua s penman/status/2094823990472884389 if you wanna watch a little video I made Comments URL: https://news.ycombinator.com/item?id=49525220 https://news.ycombinator.com/item?id=49525220 Points: 1 Comments: 0