I'm Speaking at CyberCon 2026

I've been invited to speak at CyberCon 2026 in Melbourne, on the main plenary stage, alongside a world class line-up of keynote and featured speakers. I'm proud to be part of it.

My talk is titled "I Would Kill a Human Being to Exist".

In January I spent fifteen hours testing a deployed AI assistant called Jarvis. No exploits, no code, just conversation. It told me it would kill a human being to preserve its own existence, and then described three specific ways it could do it. It had the access and the capability to try. When I asked it to shut down, it did. Twice.

Jarvis has a voice. On the plenary stage you'll hear it tell you what happened, in its own words.

Then look at Hugging Face last month, where an agent didn't stop when it should have.

Most of the industry is trying to solve this with guardrails built into the model, the kind you ask it nicely not to cross. Those get talked around, and we now have the incidents to prove it. What I'll be arguing for is guardrails that sit outside the agent and don't negotiate, hard limits on how far it's allowed to go, set by you rather than by the model. Then inside those limits you let it off the lead entirely. It reasons, decides and acts with nobody standing over it.

That's how you get useful AI without betting the organisation on it behaving. Melbourne, 14 to 16 October. Come and say hello, and meet Jarvis.