PromptDustSee what AI
leaves behind.
Cursor, Claude Code, and ChatGPT write your prompts, answers, and pasted code straight to disk — usually in plain text. PromptDust finds every copy and shows you what’s exposed.
Free and open source. Every finding stays on your machine.
The problem
The dust AI leaves behind.
It piles up on the disk: transcripts, caches, and local databases.
Every AI tool keeps records. Claude Code logs each session to a file. ChatGPT Desktop caches months of chats in a local database. Aider writes your chat history into the git repo you’re working in, as plain-text Markdown.
A pasted API key, a customer’s address, a chunk of production code. Each one lands in a transcript and sits there for months. Multiply that across a team, and every laptop carries the same buildup.
What it is
An inventory, and an exposure map.
Network tools watch the wire. PromptDust reads the disk.
PromptDust lists every AI tool on a machine and every place it writes. It reads each store, flags the ones holding API keys, personal data, or source code, and scores how exposed each one is.
Run it on your laptop to see your own footprint. Run it across a fleet, with each person’s consent, and security sees the whole estate.
How it works
One read-only pass.
- Find
Every AI tool on the machine, and every file, cache, and database it writes.
- Rate
Each store scored for what it holds: API keys, passwords, personal data, source code.
- Redact
You get the location and the type. Values stay masked on disk.
How it’s built
Read-only by design.
PromptDust opens your files read-only. It reads your disk and writes nothing back.
Findings stay on the machine that made them. The value of every secret stays masked, on the disk where PromptDust found it. Cleanup is a separate step, and you run it.
PromptDust tells you a live AWS key sits in this file. It leaves the key exactly where it lies.
Our commitments
Built into the tool.
- 01
Runs on your laptop.
Every scan happens on the machine in front of you. Findings leave only when you send them.
- 02
Opens files read-only.
PromptDust reads your disk and writes nothing back. Cleanup is a separate step you run yourself.
- 03
Masks the value.
You see the location and the type: a password, an API key, a customer record. The secret itself stays on disk.
- 04
Asks before it scans anyone else.
Scanning someone else’s machine needs their explicit, revocable consent, enforced in the tool.
- 05
Stores findings locally.
Results live on the machine that produced them. No server pools anyone’s exposure.
- 06
Shows its work.
Every score links to the evidence behind it. Ask why a file was flagged and read the reason.
- 07
Open to inspection.
The engine is open source. Read the code, and check the tamper-evident log of its privileged actions.
- 08
Stays in the AI layer.
PromptDust reads AI tools and the files they write. It stops there.
Proof, not promises
Verify it yourself.
Read the source. Check the log.
- Open core. The engine is open source. Read exactly what it does, line by line.
- Tamper-evident log. PromptDust records its own privileged actions in a log you can verify.
- Traceable scores. The scoring model is published and versioned, so you can check how any finding was rated.
- Versioned signature DB. The catalog runs on a dated signature database you can inspect.
The catalog
Where every AI tool keeps your data.
9 tools mapped so far: Cursor, Claude Code, ChatGPT, and more. Each entry lists the files, the format, the sensitivity, and how sure we are.
Run it on your own machine.
Download PromptDust, point it at your machine, and read your footprint in one pass.
PromptDust runs on your machine and reports only to you. Findings stay local. The value of every secret stays masked.
Running a fleet? Assessment across your whole org, with each person’s consent, is on the roadmap. See how it’s built →
PromptDust is an open engine today. Read it on GitHub.
Free and open source. Signature database 2026.07.2.