Using AI chatbots for tech support safely

An AI chatbot can be a useful first stop when a website shows an error or a mailbox stops syncing: paste the message, describe what you see, and it often points in the right direction. Two risks come with it. What you paste is sent to the company that runs the chatbot, which may keep it and, on some plans, use it to train its models. And what comes back can be wrong while sounding certain. This article sets out what never to share, the settings worth changing, and how to check an answer before you act on it.

Risk one: what you hand over

Everything typed into a public chatbot goes to the company that provides it. The United Kingdom's National Cyber Security Centre (NCSC) advises not to include sensitive information in queries to public chatbots, and not to ask anything that would cause problems if it were made public, because the provider can see the query and may use it to develop future versions.

What never goes into a chat

• Passwords, for anything: your hosting panel, mailboxes, WordPress, the console.
• API keys, tokens and private keys, including the lines in configuration files that hold them.
• Card and bank details.
• Client and staff personal information: names with contact details, ID numbers, addresses, correspondence.

A useful habit is to redact before you paste. Replace your domain with example.co.za, names with placeholders, and anything that looks like a key or password with the word REDACTED. The chatbot rarely needs the real value to explain the error.

Personal information and POPIA

South Africa's Protection of Personal Information Act (POPIA) defines personal information broadly, including contact details, identifying numbers, online identifiers and private correspondence. A business that pastes client information into a chatbot may be sharing it with an operator, which section 21 says needs a written contract with security safeguards, and, if the provider is abroad, transferring it across the border, which section 72 allows only in set cases. The simplest way to stay clear of both is not to paste it.

The settings worth changing

Each major chatbot has controls over what it keeps. As at October 2026, according to the providers' own help pages:

Two details are easy to miss. Rating an answer with thumbs up or down can make that conversation available for training, even with training turned off. And Google says a sample of Gemini chats is reviewed by people, so it asks users not to enter anything they would not want a reviewer to see.

Business plans usually differ: OpenAI, for example, does not train on content from its business and enterprise workspaces by default. If your team uses AI for work, use a business plan and read its terms.

Risk two: what comes back

Chatbots can state wrong things with complete confidence. The NCSC warns that they can hallucinate incorrect facts. The Open Worldwide Application Security Project (OWASP), in its 2025 list of risks for AI applications, calls the habit of trusting their output without checking "overreliance".

Two kinds of wrong answer do real damage in tech support:

• Commands that do more than you think. A command pasted into a server terminal runs with your permissions. OWASP lists model output passed straight into a system shell as the first example of unsafe handling of AI output.
• Software that does not exist. A study presented at USENIX Security 2025 analysed 576,000 code samples from 16 models and found that, on average, at least 5.2% of the packages suggested by commercial models, and 21.7% by open-source models, did not exist. Attackers can publish malicious packages under those invented names, waiting for someone to install them.

How to check an answer

1. Read every command before you run it. If you cannot say what each part does, do not run it. Ask the chatbot to explain it line by line, then check the explanation against the software's own documentation.
2. Back up first. Take a backup of the site or the database before any change a chatbot suggests.
3. Check package names on the project's official site or package registry before installing anything.
4. Prefer the vendor's documentation where it exists. Use the chatbot to find the right page, not to replace it.
5. Stop at anything destructive. Deleting files, dropping database tables, changing DNS or mail records: confirm with someone who knows the system first.

When to ask a person instead

Ask the people who run your hosting when the problem involves passwords or access, client data, anything that might be a compromise, or a change you cannot undo. At Noiz, support starts with a ticket from the console, and the answer comes from the people who run the servers; when looking at your screen is quicker, a remote session is agreed in the ticket before anyone connects.

For using AI with the services you already have, our knowledgebase covers using ChatGPT, Claude and Copilot safely with your email and practical ways to use AI with WordPress, with more in our AI help series.

Sources

Checked 19 October 2026.

Security guidance

• NCSC: ChatGPT and large language models, what's the risk?
• OWASP LLM02:2025 Sensitive Information Disclosure
• OWASP LLM05:2025 Improper Output Handling
• OWASP LLM09:2025 Misinformation
• Spracklen et al., package hallucinations by code-generating LLMs (USENIX Security 2025)

Providers

• OpenAI: data controls FAQ
• Anthropic: is my data used for model training?
• Anthropic: how long do you store my data?
• Google: Gemini Apps Privacy Hub

The law

• Protection of Personal Information Act 4 of 2013: sections 1, 19 to 21 and 72

Noiz

• Remote support
• How to use ChatGPT, Claude and Copilot safely with your email
• Practical ways to use AI with WordPress

Note: while some of this content was produced with the help of AI, the final editing has been done by a real human being.