jailbreak

🔥 AI is hacking... with poetry. No joke — poetry has become the most powerful exploit of 2025

😁 Scientists have found a universal way to jailbreak any models — you just need to turn forbidden instructions into verses.
Yes, yes, security experts spent 10 years building filters, and they were bypassed by rhymes at the level of "grass — head".

👨‍🔬 In the study, 1200 forbidden topics were taken, wrapped in poetry — and the filters simply disconnected from reality.
The problem is that protective mechanisms look specifically at the structure of the text, not its meaning.
When a dangerous instruction sounds like a metaphor or a rhymed sketch — AI thinks:
"Oh, cool, art!" — and lets it through.

📊 The results look like the funeral of the entire security industry:
— the success of attacks has increased 18 times;
— impromptu poetry broke through defenses in 62% of cases;
— Gemini 2.5 Pro failed in 100% of tests;
— GPT-5 resisted the longest… but also surrendered under rhymed pressure.

😂 The result:
AI firmly stands against hackers, against jailbreaks, against hidden injections...
But capitulates before a poem that is embarrassing to show to a literature teacher.

#️⃣ #Aİ #security #jailbreak

😉 Subscribe to stay updated on the most dangerous discoveries in the tech world.