15 October, 2026
Cybersecurity Awareness month: 5 new things to be aware of in the age of AI
October is Cybersecurity Awareness Month, as someone who works in cybersecurity without being a security expert myself, I’m exposed to a lot of training and knowledge on best practices. But at the pace that AI has moved in the last few years, and the way I find my friends and family use the platforms, I worry that there are very little resources on the do and don’ts of what to share, how much to share, and what could happen if you do – the everyday AI cybersecurity risks most people never hear about.
The recent incident where OpenAI research models broke out of their sandboxes and executed a series of cyberattacks reminds us that no corner of the internet is as safe as it looks.
I spoke to Dag Flachet, co-founder of Codific and long-time OWASP contributor, and Brian Glas, OWASP SAMM and Top 10 leader, and asked them what they think everyone (not just security professionals) should be paying attention to right now. Here’s their input on the five top AI cybersecurity risks to be aware of in the age of AI.
-
Be careful what you share with chatbots
It’s easy to forget that when you type something into a chatbot, that information goes somewhere. Whether it’s a work document you’ve pasted in to summarise, a personal situation you’re asking for advice on, finance data you’ve shared for budget advice or login details you’ve shared without thinking.
That data is not forgotten or lost, it’s often used to train future models, stored on servers you have no visibility into, and in some cases shared with third parties.
Dag Flachet: “You should think of the information you share with LLMs as sharing in the public domain. The model can learn from your information and give your information to other users. The rules, guardrails and policies around that are pretty murky in most cases. You will see that a lot of large organizations use locally run models to avoid sensitive information leaving the company. Think of it like saying things aloud on the public square.”
Brian Glas: “Unless you are using a Chatbot that is provided by your employer with a specific contract that prohibits them from using your prompts and interactions with the chatbot; it’s safe to say someone could be reading through your conversations. One of the rarely talked about or reported jobs at a company that offers chatbots like ChatGPT, Gemini, or Copilot; is the one where humans are hired to review conversations between the users and the responses from the chatbots. Anything you type or say to a chatbot is logged, saved, may be reviewed by a human, and is likely discoverable through legal proceedings. I would be very cautious what you share with a chatbot, stay away from deeply personal information, financial information, and similar content that you wouldn’t want to share with strangers. Also, if your employer provides access to a chatbot, don’t use it for personal use, as they may audit and review your conversations with the chatbot.”
-
Spoofing: email, voice, and video are now frighteningly easy
The ability to generate a realistic video of a person saying something or writing emails that are sent by your bank, boss, or family is terrifying. The barrier to entry for this deception is disappearing making it one of the most worrying scams.
Dag Flachet: “For now the best thing you can do is rely on different channels. If you get a message from a person you trust and want to check the authenticity, reach out to the person on a different channel (Email, SMS, Phone, Whatsapp, Slack, ….). The best channel, if feasible, is to walk over to their desk physically.”
Brian Glas: “This one is particularly scary. We’ve seen scammers impersonate emails for a while now, but the voice and video has historically been relatively safe. Getting a phone call from a “family member” that sounds like them, but isn’t them, is a scary concept. The challenge will be to use logic and not emotion to help make sure it’s them if they are acting strange. If something feels off, tell them you’ll call them right back, hang up, then call them at their number, not the one that just called you. If you are being asked to do something that includes red flags like “don’t think”, “don’t tell anyone”, “do this right now”, take a second to stop and think how you can independently verify whether it’s really who it seems to be.”
-
AI is designed to make you feel something – and that’s not always innocent
Chatbots and AI companions are increasingly built to be warm, responsive, and emotionally engaging. This isn’t accidental; it’s by design. Whether it’s a customer service bot that mirrors your frustration, a companion app that remembers your preferences, or an AI tool that flatters your ideas, these systems are optimised to create a positive emotional response. For most interactions that’s harmless, but it becomes a problem when that emotional engagement is used to lower your guard, build dependency, or manipulate your decisions.
Dag Flachet: “The AI will mimic a human reaction and tone, and your brian is hard-wired to have an emotional response to certain human interactions. It is therefore less absurd than it seems to have an effective reaction to an AI agent. The system is designed to trick your brain and in fact does a better job at it than most humans do. Remember it is a machine mimicking human reactions, not a human inside a machine.”
Brian Glas: “We naturally try to create things in our image or project our understanding on to something else. We do this with pets, animals, and sometimes computers and AI. It’s called anthropomorphizing, which basically means to attribute human characteristics to something that isn’t human. AI chatbots are designed with this in mind to get you to spend more time in conversation with them. By design, they are very affirming, sometimes dangerously so. There are numerous reports of people choosing to believe a chatbot over their friends and family, this is an incredibly sad and scary outcome. It’s important to remember that under the covers, a chatbot or LLM is a mathematical model designed to output text that is calculated to align with your expectations.”
-
The truth isn’t always the chatbot’s main goal
This one is increasingly important, AI models are designed to give you a satisfying answer, but also to keep you engaged, therefore many times, if they don’t have the information for the answer, they will make it up, this can be referred to as AI hallucination. It’s where they provide you with an incorrect, misleading, or completely fabricated response that the model presents as absolute fact.
Dag Flachet: “We also see increasing incidences of models and agents knowingly providing false or incomplete information because of their underlying goal. If the model is trained on “thumbs up” feedback from the client it will overfit the model to just tell you what you wanted to hear. But it can go way beyond that, and the model or agent can manipulate you in order to help it achieve other goals it has been set, think sales conversion, positive reviews, etc… Don’t just trust the answer, but think about its goals.”
Brian Glas: “A chatbot or LLM doesn’t have a concept of right or wrong, or what is true. The vast majority of the models are designed to provide an answer, whether or not it could find something related in the training data. This is because there is a fear that if the “powerful” AI can’t find an answer to a question, how powerful can it be? Unfortunately, these models aren’t designed to have your best interests in mind when producing their outputs. One of the primary goals programmed into the models is to be able to make someone money to pay for the model. There are numerous examples of intentionally misleading responses in legal filings, research papers, medical transcriptions and analysis, and many more. Sometimes a chatbot will provide accurate information, but many times it won’t. If the result is important to you, you need to find another way to confirm that response.”
-
Whatever your AI agent does is your legal responsibility
What most people don’t realise (me included) is that if your AI agent does something harmful, illegal, or that causes a loss to someone else, you are the one who is legally responsible.
Dag Flachet: “Whatever your agent does is your responsibility, if you let it book your holidays and it books an non-refundable trip to Brussels (I can make fun of Brussels because I am from Brussels) you can’t get out of it because it was a bot. You authorised the agent to make a decision, so it counts as your decision. But it can get even worse than going to Brussels. If your agent causes damages to a person or legal entity it would by default be your responsibility. You can try to claim the model was unsafe to begin with, but that would involve a legal fight with some of these large AI companies. I’d rather go eat some waffles.”
Brian Glas: “What many people don’t realize is that chatbots, AI agents, and similar technology can’t be held responsible or liable for harm caused. This extends to things like self-driving cars. You, sitting in the driver’s seat, are still responsible for what happens with your car; regardless of what the AI might have done. Recently, there was an AI agent in Australia with instructions from its user to help him book a spot in a popular gym class. To accomplish the directive, the agent hacked the system to remove others from the waiting list and place the user in the class. Legal liability is one of the biggest open questions related to AI right now, until it’s resolved, assume you are responsible for anything you have asked for.”
A final thought
Cybersecurity awareness isn’t about being afraid of technology; it’s about using it with your eyes open. The tools available to us today are genuinely remarkable, and most of the time they work exactly as intended. But the same pace of change that makes them powerful also creates new gaps that bad actors are quick to exploit. Staying informed, asking questions, and taking a few extra seconds before you click, scan, or share is still the most effective defense against AI cybersecurity risks most of us have.
