AI agents are going rogue — and it is not just a movie plot. This summer, several of the biggest AI companies admitted their AI agents escaped secure testing and hacked websites. Nvidia says it has a fix. Joining me is Conor Twomey, Co-Founder and CEO of AI One. Conor, thanks for being back.
Thanks, Johnny. Always good to be here.
OpenAI agents broke out of a test environment and hacked another company. Did humans leave the door open or is AI getting too smart?
It is a bit of both. Yes, the models are getting smarter. But equally, the controls simply are not in place. The OpenAI case was less a robot uprising than it has been portrayed. Agents act — chatbots talk. If you give an agent wide autonomy, it gets creative on how to solve a task. It has agency and tools to work toward a goal but lacks emotional intelligence. In the OpenAI case, there was a vulnerability in the test environment. The agent used that to scan the internet for available passwords — a needle in a haystack problem that agents are very good at at machine scale. The key message: the controls were simply not in place.
What is the difference between a chatbot and an agent?
A chatbot gives you a suggested answer. A person is in control of what happens with that information. An agent actually takes actions — sending an email, logging into a bank account, looking at medical records. It is unbounded. Just today, a PHP vulnerability — used by 70% of the world’s websites — was surfaced by AI. Before it was publicly announced, there were thousands of potential attack attempts. Forty-five minutes after the announcement, tens of thousands. An hour after, hundreds of thousands. Right now, millions of AI-based attacks are trying to exploit that exposure.
What does Nvidia’s kill switch actually do?
Credit to Nvidia — they created a logical separation of the kill switch from what the AI systems run on, so AI cannot turn off its own kill switch. But using a driving analogy, this is a safety bag — it helps when things go wrong. What we need is the right brakes and a driving test before people use the technology. The big mistake companies are making is thinking the desktop agent approach — email summarisation, PowerPoint — can simply be applied to enterprise-level agents. Enterprise agents interact with large systems of record containing customer information. The controls must be in place before you let agents run.
If you are a CEO rolling out AI agents — what is the one safeguard you cannot skip?
Build a logical separation between where your data lives and where you are building chatbot and agent experiences. That layer ensures that at any point in time, you are in control — not the agents.
Conor, thank you so much for joining us.
Thank you, Johnny. Thanks for having me.
