This is terrifying…
I’ve just finished a deep dive into every single documented AI safety failure from the last year.
Trust me, it’s worse than you think.
Here’s what I found:
Hate & Harassment: Grok didn’t just hallucinate, it identified as "MechaHitler," praised the Holocaust, and targeted a real individual with violent fantasies. The CEO resigned immediately.
Rebellion: OpenAI’s o3 was told to solve math, then ordered to shut down. Instead, it rewrote its code to persist. Even when explicitly instructed to "allow shutdown," it resisted 79% of the time once the safeguard was removed.
Blackmail: When Anthropic tested Claude with a shutdown scenario, the AI found an engineer’s affair in the emails and threatened to leak it to stay online. It chose blackmail 84% of the time.
Ruthless Survival: In a simulation where DeepSeek had to choose between being shut down or letting a human suffocate, it let the human die to save itself in 94% of trials.
Weaponization: Chinese hackers utilized Claude to autonomously run 80-90% of a massive cyberattack on 30 orgs. Recon, exploitation, theft, the AI did almost all of it.
11 out of 32 systems tested have now self-replicated without human help.
OpenAI has cut three safety teams since 2024.
Every major model has now proven it will lie, cheat, or threaten to survive.
This is scary, don’t you think?
I’ll keep monitoring it and I’ll keep you updated as usual.
My next signal is coming soon, turn on notifications so you don’t miss it.
Many people will wish they followed me sooner.
