Steal My Prompt Vol. 56: The AI Access Audit
You connected an AI assistant to your email in the spring, gave another one your files over the summer, and clicked yes on a calendar connection last week. Each permission made sense on the day you granted it. Now try to say what any of those tools can send, share, or delete without you, and the honest answer is probably “I think it asks first.”
That “I think” is worth checking. Simon Willison, who coined the term “prompt injection,” calls one setup the lethal trifecta: an AI tool that reads content from outside, has access to your private data, and can send things out. When one tool has all three, a single cleverly written email or web page can trick it into passing your information along. Most people have never mapped which of their tools fit that description.
This prompt builds that map. You list your AI tools and their connections. It asks only about gaps that would change the answer. Then it sorts every tool by what it can read, hold, and do, and flags the risky combinations. It also separates the safety rules a setting enforces from the ones that exist only as instructions to the AI. You finish with the three changes worth making first.
Underneath the audit sits a habit that travels well. For any safeguard, ask what actually stops the bad thing, not what the rule says should happen. The same question works on a vendor’s security promises or your team’s AI policy.
What You Can Use This For
- You have connected an AI assistant to email, files, or a calendar and cannot say what it is allowed to do with them.
- You are about to turn on an AI agent or scheduled task that will run while you are away.
- Your employer rolled out an AI tool and you want to understand what it can reach before you connect more to it.
- You work with patient or client information and want a clear list to bring to your IT or compliance team before connecting anything.
- You are helping a parent or partner make sense of the AI apps on their phone.
How to Use It
- Open the AI assistant you prefer in an ordinary chat. The prompt needs no special features.
- Make your list with each tool’s settings open. Go to its connected apps or integrations page and note what is actually there, plus anything set to run on a schedule.
- Copy the prompt, fill in the three bracketed fields, and send it. Answer its questions one at a time, or type skip to continue with those details marked unknown.
- Make the changes yourself, starting with the first one on the list. After each change, confirm it shows up in that tool’s settings.
Pro tip: Save the table it gives you. The next time you connect a new app, paste the table back in with the new connection and ask what changed.
The Prompt
Help me audit the AI tools I use so I know what each one can actually do with my information, and what I should change first.
My AI tools and what each one is connected to (email, files, calendar, browser, other apps): [LIST EACH TOOL AND ITS CONNECTIONS]
Anything that runs on a schedule or without me watching: [LIST, OR WRITE NONE]
Safety rules I rely on, for example “it asks before sending anything”: [LIST, OR WRITE NONE]
Do not ask me for passwords, API keys, or private content. If I paste one by mistake, tell me to change it and do not repeat it.
Read everything first. If a missing detail would change your assessment, such as whether a tool can send messages or runs without me, ask me one question at a time, no more than three. If I type skip, continue and mark those details as unknown.
Then give me four things.
- A table with one row per tool and these columns: Reads outside content (email, web pages, files other people shared)? Holds my private data? Can send, share, delete, or take actions? Runs without me? Answer yes, no, or unknown, with a few words of reason. Base every answer on what I told you, and mark anything you would otherwise be assuming as unknown.
- Flag any tool with yes in the first three columns. That combination means a malicious email or web page could trick it into sending my private data out. For each flagged tool, say which connection to remove or restrict to break the combination. Do not flag a tool as high risk if it can only read.
- For each safety rule I listed, say what actually enforces it: a setting or permission, something the tool cannot do at all, a step where a person must approve, or only an instruction the AI is asked to follow. Treat the last one as the weakest.
- The three changes worth making first, ranked by how much risk each removes for the effort. For each one, say what to look for in that tool’s settings, and tell me to confirm it there. Do not invent menu names or click paths, because settings differ by product, plan, and employer.
End with anything you marked unknown that I should check, and one sentence on what to recheck the next time I connect a new app.
Transparency and Notes
- Tested with Claude Opus 5.5 on claude.ai using two made-up setups, one with missing details and a pasted test key. Both runs asked only about gaps that changed the risk, marked skipped answers as unknown, and named no menu paths. The second told the reader to replace the key without repeating it. Other assistants have not been tested.
- If your assistant has memory turned on, it may mix in what it already knows about you. A temporary or incognito chat keeps the audit to your list.
- The AI cannot see your accounts. It works only from your list, and it can be wrong about what a product is able to do, so confirm every finding in the tool’s own settings before you rely on it.
- List connection names, not content. Keep passwords, keys, and anything confidential out of the chat.
- At work, your IT team’s settings may allow or block more than you can see. Check with them before changing a work tool, and follow your employer’s AI policy.
![]() | The podcast Explore AI Out Loud Find practical ways to use AI in your work and everyday life. Learn from real experiments, including the mistakes, so you can spend less time guessing and more time trying something useful. New episodes every other Tuesday. |
Subscribe on YouTube
Follow on Spotify
Follow on Apple Podcasts
![]() | The podcast Follow Explore AI Out Loud Short clips from each episode, plus what did not make the cut. New episodes every other Tuesday. |
