Welcome to the first AI security issue of The SecureByDefault Brief.
Quick promise before we start: this is not an "AI is coming for everything" newsletter. AI tools are genuinely useful, and telling people to stop using them is both unrealistic and a little silly at this point.
The real problem is quieter. Every time someone pastes something into an AI tool, that information leaves your control and lands on someone else's servers. Most of the time that is fine. Sometimes it is a customer list, a chunk of source code, or an internal document that was never supposed to leave the building.
The giveaway is not the tool. It is what you are about to hand it.
If your AI policy is "everyone just be careful," congratulations, your policy is a vibe. Let us give it some actual edges.
Real attack breakdown: the paste that got ChatGPT banned
This one is not a hacker story. It is a normal-people-doing-normal-work story, which is exactly why it is worth telling.
In April 2023, engineers at Samsung's semiconductor division used ChatGPT to help with their work. One pasted in confidential source code to help fix a bug. Another pasted internal meeting notes to turn them into a summary. Reasonable instincts. The tool is good at exactly those tasks.
The catch is where that data went. Reporting at the time described three separate incidents in about 20 days. Within weeks, Samsung restricted employee use of generative AI tools like ChatGPT, Bard, and Bing on company devices and networks. The stated reason: data sent to those services sits on external servers, which makes it hard to pull back or delete.
Here is the honest part. There is no public evidence that Samsung's code ever showed up somewhere it should not have. That is not the point. The point is that the moment it was pasted, Samsung no longer decided where it lived, who could see it, or how long it stuck around.
It is a bit like reading your diary out loud to a stranger who is very polite, takes great notes, and never forgets anything. Nothing bad has to happen for that to be a bad idea.
Why this matters right now
The security community has a name for this. In the OWASP Top 10 for LLM Applications, the 2025 edition ranks Sensitive Information Disclosure as the number two risk, up from number six two years earlier. In plain terms: people handing confidential data to AI tools is now one of the most common ways that data gets exposed. Prompt injection sits at number one, which is a topic for a future issue.
Two things make this a real workplace problem, not a theoretical one:
It does not feel like a security event. Nobody clicks a sketchy link or types a password into a fake site. They paste a spreadsheet into a helpful tool and get a helpful answer. The whole thing feels productive.
Shadow AI. Most of this happens on personal accounts and free tools that IT never approved and cannot see. If you do not give people a sanctioned way to use AI, they will use an unsanctioned one, quietly.
The pattern to remember: the danger is not that AI is evil. It is that "paste it in and see what happens" treats confidential data like it is disposable.
Practical fixes you can apply this week
For everyone
1. Run the public-channel test. Before you paste anything into an AI tool, ask one question: would I be comfortable posting this in a public Slack or Teams channel where anyone could read it? If the answer is no, stop and sanitize it first.
2. Know what is a hard no. A quick "would you paste this?" guide:
Public blog outline or general question: usually fine
Sanitized troubleshooting notes with names and IDs removed: usually fine
Customer records, employee data, or contracts: no
API keys, passwords, or access tokens: no
An internal incident report: only if your company policy and tool settings allow it
3. Sanitize instead of skipping. You rarely have to choose between "use AI" and "protect the data." Strip the names, account numbers, and secrets, then paste the shape of the problem. The tool does not need the real customer to help you write the email.
For admins and small business owners
4. Give people an approved tool. Shadow AI thrives when there is no sanctioned option. A business-tier account that does not train on your data is a far better outcome than ten employees quietly using personal free accounts.
5. Write a one-page AI policy. Not a legal document. A short, plain list of what is fine to put into AI tools, what is never fine, and which tool the company actually supports. People follow the policy they can remember.
6. Turn off training and check the account type. Make sure the default AI tools your team uses are configured so conversations are not used to train models, and understand the difference between a personal account and a business account. More on the exact settings below.
This week in "Please Don't Do That"
Someone, somewhere, is about to paste an entire customer spreadsheet into an AI tool because the columns would not line up.
Please do not hand over your whole customer list to fix a formatting problem. The tool does not need real names and account numbers to align a column. It needs the shape of the problem.
Sanitize first, paste second. Your customer list will thank you. Well, it will not, but you get the idea.
Tool worth knowing: your AI tool's own data controls
The most useful "tool" this week is a setting you already have.
In ChatGPT, personal accounts have a Data Controls section in Settings with a toggle called "Improve the model for everyone." Turn it off and your new conversations are not used to train models. There is also Temporary Chat, which is not saved to your history, is not used for training, and is deleted after 30 days.
One honest caveat, because this is exactly the kind of thing that gets oversold: turning off training does not mean nothing is stored. OpenAI can still keep data for up to 30 days for abuse monitoring on standard accounts. "It does not train on my data" and "nothing is kept" are two different promises. For business use, ChatGPT Business, Enterprise, and the API do not train on your inputs by default, which is a big part of why a business account beats a personal one at work.
If your team uses a different tool, the homework is the same: find its data controls, find out what account type you actually have, and read what the vendor itself says about retention. Not the blog posts. The vendor's own documentation.
Checklist tie-in
This issue is about a habit more than a product: know what is safe to share before you share it. That is the same instinct behind the SecureByDefault 25-Point Security Checklist, which is built around knowing where your data is and who can reach it.
One question for you
Has anyone on your team ever pasted something into an AI tool and then immediately regretted it? No judgment. Bonus points if it involved a spreadsheet, a "just this once," or a very confident intern.
Have you ever pasted something into an AI tool and immediately regretted it?
Hit reply if you want to tell me the story behind your answer. I read every one.
Talk soon,
Ron
Cloud and Cybersecurity Engineer, SecureByDefault
Sources
Bloomberg, Forbes, and CNBC reporting on Samsung restricting employee use of generative AI tools after confidential data was entered into ChatGPT (May 2023) · OWASP Top 10 for LLM Applications 2025 (LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure) · OpenAI Help Center: "How your data is used to improve model performance" and "Data Controls FAQ" · SecureByDefault 25-Point Security Checklist