Jailbreaking – a Way to Break the Ethical Safeguards of AI Models

Jailbreaks are intricately phrased AI prompts that allow powerful chatbots to sidestep the human-built guardrails governing what they can and can’t say. While they may yield dangerous information, hate speech, or simply falsehoods, jailbreaks also serve to highlight the capacity and limitations of AI models. A small but growing number of people, including anonymous Reddit users, tech workers, and university professors, are coming up with methods to poke and prod (and expose potential security holes) in popular AI tools. Crafting these prompts presents an ever-evolving challenge: a jailbreak prompt that works on one system may not work on another, and companies are constantly updating their tech.

Jailbreaks su sofisticirani AI promptovi koje omogućavaju moćnim četbotovima da zaobiđu prepreke koje regulišu šta mogu i ne mogu da kažu. Iako mogu da generišu opasne informacije, mržnju ili jednostavno laži, jailbreakovi takođe služe da istaknu kapacitet i ograničenja AI modela. Mali ali rastući broj ljudi, uključujući anonimne korisnike, tehnološke radnike i univerzitetske profesore, dolazi sa metodama za ispitivanje i prodor (i otkrivanje potencijalnih sigurnosnih propusta) u popularnim AI. Kreiranje ovih promptova predstavlja stalno rastući izazov: jailbreak prompt koji radi na jednom sistemu možda neće raditi na drugom, a kompanije stalno ažuriraju svoju tehnologiju.

Link