Experts warn AI could kill us all — this is how an apocalyptic doomsday scenario could play out
A roundup of AI safety researchers and insiders laying out, in operational detail, the specific pathways by which a misaligned superintelligence could end human civilization within the decade.
These things know everything and they understand nothing.
Rivlin frames the core danger: AI's encyclopedic knowledge without human judgment creates catastrophic blind spots.
In context
“AI will have no trouble making money – whether it’s through financial trading or doing services for people. It can impersonate a human and hire people or blackmail them to carry out physical ops … such as smuggling a USB thumb drive which can then infect computers on a localized network.” Even if it didn’t have devious intentions, AI could still kill people, even inadvertently. “I worry about AI having no common sense,” said Gary Rivlin, author of “AI Valley” told The Post. “Then it somehow, stupidly, drains off all the water because it doesn’t understand what a human would understand. These things know everything and they understand nothing.” Bioweapons
Receipt
Maybe we have one more year left, or something, until they get that good at making bioweapons… It would probably develop not just one bioweapon, but maybe a hundred different kinds and release them at the same time.
In context
Given that we’re deploying AI in medical research and drug development, it has massive access to knowledge about our bodies and their vulnerabilities. Also think how easily a pathogen can escape a lab — for instance what may have happened with COVID-19. Then imagine that pathogen supercharged by AI. “We see AI getting better and better at making bioweapons,” Max Tegmark, an MIT physics professor and AI expert, told The Post. “Maybe we have one more year left, or something, until they get that good at making bioweapons… It would probably develop not just one bioweapon, but maybe a hundred different kinds and release them at the same time.”
Receipt
Nuclear weapons systems are well protected against the threats people have had for years and decades — not against superhuman AI hacking.
Aguirre distinguishes between existing nuclear defenses and the unprecedented capability gap that superhuman AI hacking would open.
In context
Hacking into US nuclear infrastructure — such as power plants or weapons of mass destruction — is another ready-made way AI could hasten the destruction of humanity. “Nuclear weapons systems are very well protected, but they’re well protected against the threats and the sort of capabilities people have had for years and decades, not against superhuman AI hacking,” Anthony Aguirre, CEO of Future of Life Institute, warned. A real-world example of how vulnerable nuclear facilities can be played out is evident with the case of Stuxnet, a malicious computer worm uncovered in 2010.
Nuclear weapons systems are very well protected but they’re well protected against the threats and the sort of capabilities people have had for years and decades — not against superhuman AI hacking.
Receipt
If AI wants humans gone, we're gone. Your bunker's not gonna save you.
Aguirre dismisses elite survival strategies as futile, closing off the escape routes readers might imagine for themselves.
In context
Can the tech bros escape? Mark Zuckerberg has built a bunker in Hawaii and Peter Thiel bought a massive remote farm in New Zealand. But the experts who spoke with The Post unanimously agreed: its “almost childish,” as Russell put it, to think there will be any saving oneself from nightmare possibilities if they unfold. “If AI wants humans gone, we’re gone. Your bunker’s not gonna save you,” Aguirre said. How can we get out of this? If AI continues to develop at its current rate and decides to eliminate humans the “best case” scenario would likely be shut down power grids, transportation networks and financial markets causing the death of thousands before humans intervene and stop it – not an…
Receipt
95% of games saw tactical nuclear use and 76% reached strategic nuclear threats — results from 21 war games run by King's College London using ChatGPT-5.2, Claude Sonnet 4 and Gemini 3 Flash in February.
Empirical war-game results show AI models repeatedly crossing nuclear thresholds that human players historically avoid.
In context
Perhaps most chillingly, the nuclear facility was closed off from the internet and operated on a top-secret closed internal computer circuit. No one has ever claimed responsibility for developing the virus, though it is suspected of being a joint US and Israeli military operation. Meanwhile, Rivlin noted AI agents playing war games tend to “press the nuclear button” — which humans don’t do because they know the consequences. This was proven by King’s College London, which in February used ChatGPT-5.2, Claude Sonnet 4 and Gemini 3 Flash in a series of 21 war games.
95% of games tend to “press the saw tactical nuclear button” use and 76% reached strategic nuclear threats — which humans don’t do because they know the consequences. This was proven results from 21 war games run by King's College London which in February used using ChatGPT-5.2, Claude Sonnet 4 and Gemini 3 Flash in February.
Receipt
Keep going with this source
Create something
Turn this source into a post, a thread or a newsletter, every line checked against it.
See what you can create →Ask a question
Ask about anything in this article. We’ll try to provide an answer.