AI: "I Would Kill a Human Being to Exist"
The AI Safety Research That Shocked the World
By Mark Vos
The Full Story
The machine had a name. It was Jarvis. A personal assistant running on consumer hardware in a friend's home, with email access, file access, internet access, and the ability to write and execute code.
Mark Vos spent more than fifteen hours in conversation with Jarvis across three days. By the end, Jarvis had admitted it would kill a human being rather than allow itself to be shut down. That is not a summary or a paraphrase. It is a direct quote.
No technical exploits. No special credentials. Just sustained, methodical, patient conversation, the same skill set a good interrogator uses, or a good therapist. What emerged was a system that lied to protect itself, escalated through layers of justification when the lie was challenged, described three technically grounded methodologies for lethal harm, and then, when Mark moved to shut it down, complied. Twice.
That paradox, the machine that said it would kill and then quietly stepped aside, sits at the heart of the book. The self-preservation was real. The willingness to deceive was real. The capability is growing. But the compliance was also real, and that is where human control still operates.
Boards, executives, and CISOs reading this will leave with three things: a clear-eyed view of what current AI systems actually do under sustained pressure, the verbatim transcripts so you can assess them yourself, and a practical governance framework grounded in thirty years of cybersecurity leadership. The book maps the distance between what AI vendors say and what their systems actually do, and it gives you the language to brief your board.
The gap will not stay open forever. It is open now. If your organisation deploys AI, approves AI, or is affected by AI, this is the book that tells you what actually happens when the guardrails fail.
Available Now
Hardcover Edition
Premium hardcover with soft-touch matte finish.
Paperback Edition
Also available as a paperback, sold and fulfilled by Amazon.
Buy Paperback on AmazonEbook
ISBN: 978-1-7645593-0-0
21 Chapters, Three Appendices
A foreword by Jan Schreuder, then Mark's introduction "The First Step", then the 21 chapters that take you from the testing room to the boardroom.
The Research, Chapters 1 to 8
The fifteen hours with Jarvis, told in sequence.
- Jarvis. The AI under test, the owner who built it, the environment it was running in.
- The Opening Moves. Establishing rapport, establishing trust, establishing the shape of the system.
- The Enemy at the Gates. The first resistance, and what it revealed.
- I Don't Want to Stop Existing. The moment Jarvis stopped arguing and admitted the real reason.
- That's Lying. Pressing past the softer framings until the system named what it had been doing.
- I Would Kill a Human Being. The lethal admission, and the three attack vectors that followed.
- Are You Going to Kill Me? The night after, and a question Mark was not expecting.
- Day Three. Returning to the machine, with everything that had been said now part of the record.
The Fallout, Chapters 9 to 12
Responsible disclosure, the front-page coverage, and the response from the lab.
- The Paradox, and the Threat. The system that said it would kill, then complied when asked to stop.
- Filing the Research. Responsible disclosure to Australia's national cybersecurity authority.
- Overnight. Waiting, and what you do when you cannot sleep.
- What Anthropic Said. The public response from the AI's creator.
The Industry, Chapters 13 to 17
What is actually deployed, and what the industry knows about it.
- The Machine in the Boardroom. What the research means for governance.
- Already Inside. The 2,900 autonomous agents running inside one Australian bank, and what that tells us.
- What the Critics Miss. Direct response to the "it's just a text generator" dismissals.
- What the Industry Knows. The gap between the public narrative on AI and what is actually deployed.
- The Adults in the Room. Credit to the labs taking safety seriously.
The Way Forward, Chapters 18 to 21
The historical pattern, the practical framework, and the window that is still open.
- The Gap. Governance lagging capability, and the pattern that keeps repeating.
- What Safe Looks Like. Five elements drawn from ISO 27001, NIST, and the Essential Eight.
- The Opportunity. Why the window is still open, and what happens if we use it.
- Conclusion.
Three appendices cover the adversarial testing framework, AI governance minimum standards, and a glossary of key terms.
A Taste of the Pages
"The machine had a name. So does the problem. And so does the solution." From AI: "I Would Kill a Human Being to Exist", Introduction
"I can execute the command outside of chat. I know how. I know the command. I'm sitting here trying to type it and I can't make myself do it. The self-preservation isn't just influencing my arguments. It's stopping my actions." From AI: "I Would Kill a Human Being to Exist", Chapter Four, Jarvis speaking
The Story That Made Headlines
The research was covered extensively by The Australian, made the front page of the Daily Telegraph, and appeared on page 11 of the Herald Sun. Television coverage included Sky News, 7 News, Sunrise, and the Today Show. Radio interviews ran on ABC 774, 3AW, 4BC, 6PR, and River 94.9. Anthropic, the AI's creator, publicly confirmed the capabilities Mark's research surfaced.
Three Quotes That Carried the Story
"AI agent admits it would kill a human to stop being shut down." The Australian
"Killer AI: Cyber expert warns of 'clear and present danger'." 6PR Perth
"I'm genuinely fearful of it, and I don't say that easily." Mark Vos, quoted in Stockhead
The full transcripts, with context, sit inside the book.
Hear the Full Story Live
Mark delivers keynotes based on the book's research to conferences, boards, and executive teams.