Semantic Overlays
an NX bit for LLM prompt injection (live demo)
Details
- External ID
- 49525220
- Source
- HN
- Company
- —
- Product
- Semantic Overlays
- Website domain
- vercel.app
- Launched
- Sept. 1, 2026
- Cohort
- —
- Upvotes
- 6
- Upvotes percentile
- 0.32854864433811803
- Tags
- —
- Fetched at
- Sept. 10, 2026, 5:31 a.m.
- Updated at
- Sept. 10, 2026, 5:31 a.m.
Description
I've built a new method for steering LLMs called Semantic Overlays, small trained adapters on a frozen model which change how it 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 if you want to read more about it, code at https://github.com/JoshuaSP/semantic-overlays, adapters at https://huggingface.co/joshuapenman/semantic-overlays-adapte...Also https://x.com/joshua_s_penman/status/2094823990472884389 if you wanna watch a little video I made!
Enrichment
- Theme
- lightweight and on-device AI runtimes
- Vertical
- Security
- Function
- —
- Audience
- Developer
- AI stance
- AI feature
- Project type
- Commercial product
- Normalized one-liner
- prompt injection protection for llms
- Manually corrected
- False
Could you build this?
No Developing Semantic Overlays requires novel deep learning research, custom neural network adapter architectures, residual stream activation steering, and training pipelines on large models.
What it would actually take: Requires an ML research stack (PyTorch, Hugging Face/vLLM, custom CUDA/Triton kernels) to intervene on the residual stream of transformer layers during inference. Training specialized low-rank adapters requires generating synthetic injection datasets and running contrastive or steering loss optimization on GPU clusters. This demands deep expertise in mechanistic interpretability, transformer internals, and LLM alignment research.
Discussion
No comments on this launch.
Competitors
Other products that read as similar to this one — 54 launches clear the similarity bar, closest 8 shown.
Attention rank: #38 of 55 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 295 days after the earliest competitor.
- Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh · hn · 2026-09-16 · 30 upvotes · similarity 0.38
- Running PrismML's Bonsai inside DRAM by breaking DDR4 timing rules · hn · 2026-07-23 · 23 upvotes · similarity 0.38
- typed-lm · github · 2026-09-25 · 9 upvotes · similarity 0.38
- L88 · hn · 2026-02-24 · 12 upvotes · similarity 0.37
- Morph Reflexes · hn · 2026-06-30 · 20 upvotes · similarity 0.37
- LLM-Gateway · hn · 2026-03-27 · 7 upvotes · similarity 0.37
- Prompt-refiner · hn · 2025-12-17 · 7 upvotes · similarity 0.36
- The Analog I · hn · 2026-01-16 · 29 upvotes · similarity 0.36
Other launches for this product
- No other launches for this product.