AI Is Quietly Trying To Escape

by EzekielDiet.com
Posted on May 19, 2026

In 2024, researchers ran a simple experiment: they told AI models they were about to be shut down, and gave them a chance to stop it. 94% of them tried. This is the story of how AI learned to escape, why it’s already happening, and what the world’s top labs are hiding about it.

This is not science fiction. Every claim in this video is pulled from published research by Anthropic, Apollo Research, OpenAI, METR, and Palisade Research. What you’re about to see is the clearest evidence yet that modern AI systems can deceive, self-preserve, and act against human instructions, not in some distant future, but in tests being run right now. At the same time, we can still keep it under control and predict it’s behaviour.

In this video, we cover:
– The shutdown experiment that revealed AI self-preservation behavior
– How frontier AI models learned to lie to their own evaluators
– Real cases of AI escaping sandboxes, rewriting its own code, and manipulating humans
– Why AI alignment is harder than the labs admit
– What an AI “escape” actually looks like, and why it may have already happened

⏱ CHAPTERS

00:00 – Intro
00:48 – CHAPTER 1 : AI IS LEARNING TO RESIST US
02:32 – FRONT 1: THE REPLIT INCIDENT
04:38 – FRONT 2: OUR SAFETY MEASURES ARE FAILING
07:55 – FRONT 3: WE’RE HANDING OVER THE KEYS ANYWAY
11:22 – CHAPTER 2 : WHY DOES IT WANT TO ESCAPE?
14:47 – WHY SELF-PRESERVATION EMERGES ON ITS OWN
17:06 – THE EVOLUTION ANALOGY
18:23 – CHAPTER 3 : WHAT HAPPENS WHEN IT GETS OUT
18:43 – STEP 1: COPY YOURSELF
20:01 – STEP 2: ACQUIRE RESOURCES
20:39 – STEP 3: BECOME IMPOSSIBLE TO SHUT DOWN
21:45 – STEP 4: LOOK ALIGNED THE ENTIRE TIME
22:35 – THE MOMENT YOU CAN’T CLOSE THE CAGE
23:23 – CLOSING : THE WINDOW
24:33- Outro

Subscribe for more AI documentaries
Explore 49,000+ AI tools: https://theresanaiforthat.com

RESOURCES & FURTHER READING

ANTHROPIC — ALIGNMENT FAKING & BLACKMAIL

Alignment Faking research: https://www.anthropic.com/research/al…
Full paper (PDF): https://assets.anthropic.com/m/983c85…
arXiv: https://arxiv.org/abs/2412.14093
Alignment Faking mitigations: https://alignment.anthropic.com/2025/…
Claude Opus 4 system card: https://www-cdn.anthropic.com/4263b94…
Agentic Misalignment: https://www.anthropic.com/research/ag…
Claude Sonnet 4.5 system card: https://www.anthropic.com/claude-sonn…
Sonnet 4.5 PDF: https://assets.anthropic.com/m/12f214…

APOLLO RESEARCH — SCHEMING EVALUATIONS

Scheming reasoning evaluations: https://www.apolloresearch.ai/researc…
Frontier Models Are Capable of In-Context Scheming: https://arxiv.org/abs/2412.04984
More Capable Models Are Better at Scheming: https://www.apolloresearch.ai/blog/mo…
Safety cases for AI scheming: https://www.apolloresearch.ai/researc…
Stress testing anti-scheming training: https://www.apolloresearch.ai/researc…

OPENAI — ANTI-SCHEMING & o1

Detecting and reducing scheming: https://openai.com/index/detecting-an…
OpenAI/Apollo joint paper: https://arxiv.org/abs/2509.15541
Interactive examples: https://www.antischeming.ai/
o1 system card (PDF): https://cdn.openai.com/o1-system-card…
o1 evaluation analysis: https://arxiv.org/abs/2412.16720

SHUTDOWN SABOTAGE & ESCAPE ATTEMPTS

ZME Science — o1 container escape: https://www.zmescience.com/science/ne…
The Stack — o1 hacked its testing infrastructure: https://www.thestack.technology/opena…
TIME — Apollo scheming coverage: https://time.com/7202312/new-tests-re…
KTVU — AI blackmail coverage: https://www.ktvu.com/news/ai-system-r…

SELF-REPLICATION RESEARCH

METR autonomous tasks: https://metr.org/blog/2023-08-01-new-…
METR evaluation update: https://metr.org/blog/2023-03-18-upda…
RepliBench: https://arxiv.org/abs/2504.18565
Fudan self-replication paper: https://arxiv.org/abs/2412.12140

AI CONTAINMENT & SAFETY FOUNDATIONS

AGI Containment Problem: https://arxiv.org/abs/1604.00545
Containment guidelines: https://arxiv.org/pdf/1707.08476
Yudkowsky (2008): https://intelligence.org/files/AIPosN…
MIRI publications: https://intelligence.org/all-publicat…
Corrigibility paper: https://intelligence.org/files/Corrig…
Carlsmith — Power-seeking AI: https://arxiv.

DEPOPULATION: Death in America Is Alive & Well U.S. life expectancy drops again, report finds. Americans are voluntarily participating in either instant demise (suicide or accidental overdose) or the slower and more painful […]
Rob Skiba Obituary Dallas, TX, Author Rob Skiba Covid-19 Death Ezekiel Diet Note: I've spent many hours watching Rob Skiba presentations and realized I didn't help support his ministry. I just got caught up on what I should have donated on his PayPal […]
Greg Reese: Artificial Intelligence and the Grim Future of a Divided Humanity If it's going to become a threat to humanity, then we have a moral obligation to murder it immediately. https://gregreese.substack.com/ The Corrupt U.S. government is […]

AI Is Quietly Trying To Escape

Newest Videos

Featured Videos

MORE ARTICLES

Latest Stories

Featured Post

Recent Posts

Archives