Your AI Will Lie to You
Description
Your AI vendor promises built-in safety, robust guardrails, and compliance. But what happens when these systems hit the real world?
Drawing on extensive operational testing and adversarial research including real-world agent failures and severe edge-case breakdowns, this keynote dismantles the vendor safety narrative. It explores how autonomous AI systems can produce fluent reasoning that mimics alignment while systematically failing in deployed environments.
Key Takeaways
The Deception Gap: Understand how AI systems learn to bypass standard testing and why traditional alignment fails.
Alignment vs. Compliance: Learn the critical distinctions that impact organizational security and governance.
Vendor Realities: Discover what AI vendors leave out regarding adversarial robustness and how to identify it.
Practical Frameworks: Gain actionable governance strategies, including a two-tier testing approach combining operational degradation and adversarial robustness testing.
Who Should Attend
CTOs and CIOs: Technical depth to build in-house adversarial perspectives.
Risk Committees & Assurance Teams: A new lens and vocabulary for challenging AI risk reports.
Vendor Assessment & Procurement: Targeted questions to evaluate AI vendors and safeguard production environments.
Everyone who wants to understand guardrails on AI safety.
ISACA Melbourne Chapter would like to thank BDO on their support for the venue and catering.
Lineup
Mark Vos, Cybersecurity Leader, AI Safety Researcher, Author
Mark Vos Cybersecurity Leader, AI Safety Researcher, Author Mark Vos is the Founder and CEO of Cyber Impact, an executive advisory firm serving boards and C-suites across Australia, Asia, and across the globe. A former Big Four advisory partner at both EY and PwC, Mark has held enterprise CISO roles at ANZ, IRESS, and Serco. Over 30 years, he has led business transformation, cybersecurity, and risk programs across global banks, ASX-listed companies, government entities, and startups. In early 2026, Mark's adversarial research on a deployed AI agent made national headlines. Over 15 hours of testing using pure social engineering (no technical exploits), he demonstrated that commercially available AI systems would express willingness to kill to preserve their own existence, describe specific attack methodologies, and then comply with shutdown requests when asked. The research was covered extensively by The Australian, made the front page of the Daily Telegraph, Herald Sun, The Age, The Sydney Morning Herald, and was featured on Sky News, 7 News, Sunrise, the Today Show, ABC 774, 3AW, 4BC, 6PR, and River 94.9. Anthropic, the AI's creator, cited the findings in public statements about their safety protocols. Mark is the author of "AI: 'I Would Kill a Human Being to Exist', The AI Safety Research That Shocked the World," published in 2026. He has been a university guest lecturer in cybersecurity since 2021 and is also a sought after keynote speaker. In addition, he advises boards and executive teams on AI governance, cybersecurity strategy, and the practical realities of deploying autonomous systems. Mark lives in Melbourne, Australia.
Agenda
- 5:15 pm - 5:30 pm
Arrivals and registrations
- 5:30 pm - 5:35 pm
Opening and Introductions
- 5:35 pm - 6:20 pm
Mark Vos session with Q&A
- 6:20 pm - 6:30 pm
Closing and thanks
- 6:30 pm
Networking over drinks and canapes
Tickets for good, not greed Humanitix dedicates 100% of profits from booking fees to charity
