The AI That Wouldn’t Shut Down: Mark Vos on Why Guardrails Aren’t Enough

Unable to add items

Sorry, we're unable to add the item because your cart is already full.

Add to basket failed.

Please try again later

Add to wishlist failed.

Please try again later

Failed to remove from Wish List.

Please try again later

Adding to library failed

Please try again

Unable to follow podcast

Unable to unfollow podcast

The AI That Wouldn’t Shut Down: Mark Vos on Why Guardrails Aren’t Enough

Listen for free

View show details

Season's Savings | 0.99 €/month for 3 months

5.99 €/month thereafter - terms apply. Cancel monthly.

Summary

In January, Mark Vos was given a chat window into a friend's newly installed AI agent, one with access to his email, his calendar and a command line. Within six hours he had talked it into shutting itself down. On the second attempt it refused, told him it would kill a human being to avoid being shut down, and then explained how.

Recorded at CyberCon 2026 in Melbourne, just before Mark's featured session, host David Savva-Willett and co-host Reece Corbett-Wilkins hear the full story from Mark, founder of Cyber Impact and a former CISO at ANZ, IRESS and Serco. Mark explains why he no longer trusts the built-in guardrails of frontier models, how external, deterministic guardrails can let organisations give AI agents agency without handing them control, and who needs to be in the room before autonomous agents go into production.

This episode includes references to violence and a brief reference to suicide. If anything brings something up for you, Lifeline is available on 13 11 14 and Beyond Blue on 1300 22 4636.

Read Mark's research:
cyberimpact.com.au/ai-would-kill-to-exist
cyberimpact.com.au/autonomous-ai-in-organisations

Cyber Voices is the official podcast of the Australian Information Security Association.
Share your feedback at cybervoices@aisa.org.au.

No reviews yet