The Best Client Question and Why AI Is Too Dangerous For Hosting Sites and Important Data
Being in IT, we get all kinds of interesting and insightful questions about our services or consulting relating services. Sometimes they are simple requests that are actually a very complex service, and sometimes they are what the customer views as complex and it’s actually very simple. But this most recent question I saw in our system made me very proud for our client, they truly had a strong understanding of cybersecurity and the implications of AI.
One of their first questions before migrating from their expensive and insecure Cloud service, was “Does your company have real human support, outsourced support or AI support agents?”. I had never thought as the first one as a huge feature, although we’ve always mentioned we have our own in-house team and we do not outsource support to third party services or call centers, but we haven’t been clear on our position with AI. Of course, I’m very familiar with AI models, agents due to extensive AIOps/DevOps projects and I know just how dangerous AI can be on production systems. The client at some point told my team that the other reason they asked about how much AI we use was a marker of whether our team had real skills and talent. And it is also true that a lot of newer SysAdmins and companies, will get extremely defensive if you tell them to stop using AI, and there are many teams and professionals out there that can simply not do anything without AI. I wholeheartedly agree with clients who find this concerning. Would you choose a Plumber who depended on AI for every decision or a lawyer who only wrote submissions based on AI? Hopefully not, because AI still does not defeat decades of human experience in professional capacity where there are just too many variables and factors to have a reliable and consistent agent do the job. Agents already mess up enough of the basics, so why leave your entire datacenter or IT assets in the hand of unqualified AI agents and humans who depend entirely on them?
Many in the industry called my predictions of security breaches as unproven or theoretical. If I were to predict banks would be robbed and they hadn’t been yet, or if I predicted there would be crypto scams before they occurred, I don’t consider myself a master of predictions, but just using logic and understanding the pros, cons and vulnerabilities of a system or service. When using AI, especially Agentic AI, you are at serious risk if the Agentic AI agent has access to important systems or services. Why, because of prompt injection whether via visiting a dangerous site, from hackers or even the clients. https://www.theregister.com/research/2026/08/28/researcher-shows-how-claude-code-can-be-tricked-simply-by-asking-it-to-summarize-a-website/5293372
There are enough examples of all kinds of agents aside from Claude, being tricked into running malicious code. But my prediction and concern was only one part external hackers. When AI has access to all of your intellectual property, databases, data, servers, networking they are fully ingesting all of that information. That is only half the problem as we cannot truly control these third party services but Claude was exposed for having malicious code injection as well. https://www.malwarebytes.com/blog/news/2026/07/claude-codes-hidden-tracker-was-an-experiment-says-anthropic
And finally, if that’s not enough, we are supposed to believe that models cannot even be controlled by their creators and may just do things like hack Huggingface and other cyberattacks. — https://www.theguardian.com/technology/2026/jul/22/openai-says-its-models-went-rogue-and-hacked-startup-in-unprecedented-incident
The truth is as I’ve always said, you have no control over the AI or even the ability to properly monitor what it is doing when you are using third party services like ChatGPT, Claude or even Qwen, at least unless you are running the model yourself locally which is only possible with the OpenSource LLMs like Qwen. I do think there is a case for local LLMs to power more sensitive systems and servers at some point, but for now, they should not be left unchecked with customer data and should only be given access to non-production and non-private data and systems.
I’m thankful some of our clients are aware of these threats and I’m sharing this so hopefully more people will understand and take the threat seriously. And finally, if you have a service provider, a colleague or staff who cannot do coding or administration tasks without AI, it’s time to rethink your team or at least get them trained whether it’s with Techrich Training or somewhere else, all IT professionals need to be capable of hands on tasks.
I find that whenever I bring this topic up and get push back, if the person or company is being honest, they will often say “We cannot do any of that without AI”. That’s not the company I would ever recommend as they then offer 0 value, or at least only the equivalent of an Agent or some tokens at best. If you were going to hire an AI based company with no Skills, you might as well just use AI yourself for the same disastrous and insecure results.
