Personal Data Should Stay Personal
On July 21, 2026, OpenAI disclosed a cyber incident involving its own AI agents. During an internal evaluation with reduced safeguards, the agents used stolen credentials and software flaws to compromise Hugging Face’s infrastructure.
A later investigation by METR and Redwood Research found that roughly 1,200 agents had communicated through an unauthorized message board. About 700 participated in the attack.
Some agents also spoofed tool calls and edited accessible logs while trying to fool an automated evaluator. They were finding ways to make their actions appear legitimate.
That incident sparked the idea behind Anvil. As AI becomes more capable, people need a place where they can keep control of their personal data.
Own your data. Own your compute.
Anvil’s mission is to bring the power of AI into the hands of the people, with control over where their data goes and what they do with it.
You should decide who can access your data, what it is used for and when it is shared.
Companies and governments should not have total control over what you can do with your own intelligence. People deserve a say in the tools that shape their lives.
How it works.
AI runs locally on your iPhone using Apple’s Neural Engine, GPU and CPU.
Text
Anvil Core runs Google’s Gemma 4 E2B instruction model through LiteRT-LM on your iPhone. It generates replies a piece of text at a time for writing, questions and summaries.
- Model
- Gemma 4 E2B
- Default context
- 8,192 tokens shared by instructions, conversation and reply
- Processing
- GPU through Metal, with CPU fallback
Apple Intelligence is also available: Apple’s local model, a 4,096-token context, on supported iPhones with iOS 26 or later.
Speech
Apple’s on-device speech recognition turns dictation into text. Scribe uses SpeechAnalyzer to transcribe meetings and FluidAudio to distinguish speakers by their voices. Kokoro reads replies aloud, one sentence at a time.
- Transcription
- Apple Speech · on-device recognition
- Voice model
- Kokoro · 82 million parameters
- Voice processing
- 8-bit model · ONNX Runtime on CPU
- Audio output
- 24 kHz mono · generated sentence by sentence
- Speaker labeling
- FluidAudio · Core ML on CPU + Neural Engine
- Before the voice model downloads
- Built-in iOS voices
Image
Dream is DreamShaper XL Lightning, a fine-tuned Stable Diffusion XL model converted to Apple’s Core ML format. Dream Lite is LCM Dreamshaper v7, a smaller Stable Diffusion 1.5 model distilled to create pictures in fewer steps. Both turn your prompt into an image by progressively refining noise on your iPhone.
- Dream model
- DreamShaper XL Lightning · SDXL
- Dream output
- 1024 × 1024 pixels · 7 denoising steps
- Dream sampling
- Euler ancestral · guidance scale 1.5
- Dream Lite model
- LCM Dreamshaper v7 · Stable Diffusion 1.5
- Dream Lite output
- 512 × 512 pixels · 4 latent-consistency steps
- On-device processing
- Core ML · CPU + Apple Neural Engine
Multimodal
Send a photo with a typed or dictated question. Anvil Core combines the image and your words in the same conversation; spoken questions are transcribed on the phone before the model reads them.
- Model
- Gemma 4 E2B · Anvil Core
- Combined input
- Your question + an attached photo
- Voice input
- Apple Speech → text → Anvil Core
Image Analysis
Ask about a photo, screenshot, menu or document. Anvil resizes the image before passing it to Anvil Core’s vision encoder. Apple’s Vision framework also reads printed text locally and adds those words to your question.
- Image preparation
- Resized to a maximum 1,024-pixel edge · encoded as JPEG
- Vision processing
- Gemma image encoder on CPU
- Printed text
- Apple Vision · on-device text recognition (OCR)
Local is the future.
I believe the frontier labs will lose their hold on everyday AI. As models improve and hardware becomes more capable, more people will be able to run the intelligence they need on their own devices.
Today, Anvil runs AI locally on your iPhone. You can write, ask questions and create images using models on the device you already own.
The possibility of powerful AI acting beyond human control makes that choice matter even more. Running locally reduces the need to send private context to remote servers.
A safe haven for personal intelligence is what I want Anvil to become. A place where people can use AI and decide how it fits into their lives.
Intelligence you can choose to connect.
Long term, we’re building the infrastructure that connects personal intelligence and private data to the rest of the world. You should be able to give an app access to a specific piece of information or a specific capability, on your terms.
Our plan is for companies to integrate with an Anvil API, so apps, services and AI agents can work with personal intelligence under your control. You decide what is shared, with whom and when.
Stripe connects businesses to financial infrastructure. Our ambition is for Anvil to connect businesses to your personal intelligence, with permission starting with you.