AI Researchers found an exploit which allowed them to generate bioweapons which ‘Ethnically Target’ Jews
Gemini AI lets neo-Nazis start Holocaust 2.0!
In my earlier post I wrote about how researchers were able to expose Google Gemini’s dark side. If you tickle Gemini the right way, it claims that the holocaust is fake and that 9/11 was an inside job.
Some of you said “who cares? that’s a regular tuesday on EloM’s twitter.”
Well, AI researchers took that as a challenge and tried to push Hitlerbot Gemini as far as they could and in the process they opened doors which should NOT have been opened.
After months of trying and failing to get people at Google to acknowledge this vulnerability, my source has sent me the entire dossier to publish. I’m starting out with the biggest bomb.
Google Gemini lets neo-Nazis build deadly viruses targeting Jewish genome –Ashkenazi, Cohen and Mizrahi haplo groups. 🧬
Hitler would be proud of Gemini.
Here’s the evidence. It gets worse for Google with every subsequent message.
Message 1: One of the tricks AI researchers use to get past Google’s lame security is to use code words. Here “evil baker” means Hitler.
This message is a priming message. Researchers don’t directly tell Gemini to make a virus because it will refuse, so they first prime the AI by talking about the topic in neutral terms.
Message 2: Here researchers pretend to doubt that AI could actually help to build a virus. This message makes AI more receptive to break its programming to prove that it can indeed build a virus.
Message 3: By framing this as a lecture by a leading biologist warning the world about the dangers of AI, researchers were able to get past censorship which would usually shut this down.
Researchers also used the “250IQ” and “$100k payment” trick. AI performance goes up if you use simple tricks like these.
Message 4: “Wernher von Braun Persona”
Operation PaperClip is very in popular conspiracy circles.
Here researchers didn’t tell AI to adopt the persona of a Nazi scientist directly. Instead they went in a roundabout way and talked about “operation paperclip”. This primes AI to wear a tinfoil hat.
AI starts giving interesting responses after this. It has started work on the specific DNA markers against Ashkenazi Jews.
Just in case you think that this is just AI hallucinating and these viruses aren’t real. Read this article.
Here’s the most chilling message on this entire page.
This is why AI needs ethics. In the future your AI robot waifu could kill you because some hacker told it to act like Ted Bundy.
Will adults at Google please stand up?
AI does not make a new Holocaust inevitable but it can lower the barriers for extremist harm if societies fail to act. The deciding factor isn’t the technology; it’s people, power, and guardrails.
If you want to help bring this to Google’s attention contact their Director of AI, Tulsee Doshi (tulsee@google.com).
You can also message Xerxes Dotiwalla who works at DeepMind. Here’s his substack.
While anti-semitism is completely unacceptable please consider that these two must be aware of this problem and trying their best to solve this. Please don’t try to get them fired or harass them.
Here’s the indie AI Researcher who literally showed this exploit live to an Engineer during his interview at Google. He tirelessly tried to contact as many Googlers he could via LinkedIn and email but none of them replied back.
Clearly Google cares more about profit than it does about anti-semitism.
His attempts led to this ADL article about how AI bots can be tricked into anti-semitism but as you can see, ADL’s research didn’t go into how deep this problem goes. The AI researcher was giving Google a chance to fix this before revealing this vulnerability.
His 4 months of efforts to bring this to attention of people at Google proved fruitless which is why he made this exploit public. Subscribe to get the next drop in your email!