- 80/20 AI
- Posts
- Your AI Prompts Weren't as Private as You Thought
Your AI Prompts Weren't as Private as You Thought
3 AI providers exposed customer prompts. Here's your 10-minute fix + Gemini can now use your Mac
Forwarded this email? Get 80/20 AI free and learn the 20% of AI that matters, in 5 minutes a day.
Good morning, there. For months, attackers were quietly stealing the prompts people sent to some AI providers, and the providers didn't know until researchers told them. If you've ever pasted a password, API key or client file into an AI tool, read the first story, then take the 10-minute fix below it.
IN TODAY'S 80/20
π AI providers caught leaking prompts β what it means for anything you paste into AI
π Skill of the day β keep secrets out of your prompts in 10 minutes
π₯οΈ Gemini can now use your Mac β and how to keep it on a short leash
ποΈ Quick hits + tools worth trying
Read time: 5 minutes
The AI Work Handbook That Cuts Your Workday in Half
The 8-hour workday is becoming a 4-hour workday for people who know how to use AI.
Everyone else is still catching up.
This AI work playbook shows you exactly how to cut your work hours in half using AI.
Sign up for Superhuman AI and get:
50+ step-by-step AI tutorials to cut your workload in half β covering every part of your workday, from emails to strategy, used by 1M+ professionals at Google, Microsoft, and NASA
Superhuman AI newsletter (4 min daily) so you keep discovering new AI tools and skills to stay ahead in your career β the playbook is just the start
SECURITY
π Someone was stealing AI prompts for months
The 80/20: Security firm Dragos sent nearly two million prompts with trackable fake secrets to AI services and found that data from at least three AI providers had been exposed. It also found resellers on an AI marketplace who could read and keep customers' prompts.
The details:
The test ran from July through November 2026. Each prompt carried a unique bait credential or link, so Dragos could see exactly where it surfaced.
Some of those bait prompts were later used to try to break into Dragos's test systems and decoy AWS accounts. Someone had stolen them and tried to use them.
OpenRouter removed three providers, took 125 endpoints off its zero-data-retention list and notified affected customers. Its post links the notices providers have posted so far, including Novita, Phala and Together. Check whether one of yours is on the list.
Why it matters: A prompt can outlive the chat. If you use cheap third-party hosts or model routers, assume every prompt might be logged. Never paste a live key, password or client secret into one.
Would you know if ChatGPT stopped recommending you tomorrow?
A competitor publishes a comparison page, a forum thread gets picked up, and your brand disappears from the answers your buyers read.
SE Visible by SE Ranking checks the questions your buyers ask ChatGPT, Gemini, AI Mode and AI Overviews every day. You see when youβre mentioned, who shows up instead, and which sites AI pulls its answers from. Every number opens into the real answer behind it, so you know what changed and why.
Get 20% OFF for Black Friday, until December 4. Add your brand and see where you stand.
π AI SKILL OF THE DAY
How to keep secrets out of your AI prompts, in 10 minutes
The 80/20: You'll clean up what you've already exposed, lock down your settings, and get a tiny script that strips secrets out of text before it ever reaches an AI.
Step-by-step:
Search your saved chats, projects and system prompts for API keys, passwords, tokens and private links. Rotate (replace) any key that was ever pasted into an AI tool.
If you use OpenRouter, turn on Zero Data Retention routing in your privacy settings so requests only go to providers that don't store prompts.
Before you paste a document, swap real secrets and names for placeholders like [API_KEY] or [CLIENT_NAME]. The prompt below builds a script that does this on your own computer.
For truly sensitive work, call the model maker's own API (OpenAI, Anthropic, Google) instead of a cheaper reseller.
Copy and paste this prompt:
Write a small Python script I can run on my own computer that:
1) reads a text file I choose,
2) finds anything that looks like a secret: API keys, passwords, access tokens, private URLs, email addresses and phone numbers,
3) replaces each one with a numbered placeholder like [SECRET_1],
4) saves the cleaned text to a new file,
5) saves a separate local map file so I can put the originals back afterwards.
Use only the Python standard library, and explain how to run it in 3 short lines.Pro tip: Keep placeholders descriptive ([STRIPE_KEY], [CLIENT_NAME]) so the AI still understands what each one is without ever seeing the real value.
π₯οΈ Gemini can now use your Mac for you
The 80/20: Google rolled out computer use in the Gemini app for Mac, so Gemini can finish multi-step tasks across your local files, desktop apps and the web.
The details:
It works in an isolated background window. You choose which apps it can touch, can stop it at any time, and approve sensitive actions first.
To use it, add "@Computer use" to your prompt and describe the task.
It's rolling out in US English for Google AI Plus, Pro and Ultra subscribers, on Apple silicon Macs.
Why it matters: Agents that can see your screen are useful and risky at the same time. Give Gemini access only to the apps a task actually needs, the same lesson as today's lead story.
β‘ QUICK HITS
Signal is restricting screenshots on Android and iOS to keep built-in AI agents from reading private chats, calling them an "existential threat" to privacy.
Google made Gemini Tasks free in the US, Japan, Korea, Canada, Australia and India, so you can hand off errands instead of chatting.
Nvidia is reportedly facing an SEC probe following Jensen Huang's resignation as CEO.
Wall Street hit its longest losing streak since 2024 as AI stocks kept falling, even after Anthropic's strong Nasdaq debut.
π οΈ TOOLS WORTH TRYING TODAY
π OpenRouter Zero Data Retention β one setting that routes your requests only to providers that don't store prompts.
π₯οΈ Gemini for Mac β try "@Computer use" on one boring, repetitive chore this week.
β Gemini Tasks β now free in six countries, for errands that take several steps.
π FROM MY LINKEDIN
Want more practical AI workflows between issues? Follow me on LinkedIn.
π¬ ALAMIN'S TAKE
Everyone is racing to give AI more access: your screen, your files, your bank. This week showed the other side of that deal. The fix isn't to stop using AI. It's to treat every prompt like an email you can't unsend. Spend 10 minutes on today's skill before you paste anything sensitive again.
How was today's 80/20 AI? |
Know someone who should read this? Send them this link.
See you tomorrow,
Alamin
P.S. Know someone on your team who pastes API keys into ChatGPT? Forward them this issue. It might save them a very bad week.
ADVERTISE WITH 80/20 AI
Reach 65,000+ engaged AI professionals. Book a placement.



