A jailbroken model is a generative model whose safeguards or content restrictions have been bypassed, removed, or deliberately avoided, allowing outputs that mainstream systems would normally refuse. Some are modified versions of existing models; others are marketed as ready-made tools for abuse. section 5.3
The main significance lies less in novel capabilities than in easier access to harmful ones:
- Models advertised as uncensored alternatives may generate phishing emails, malicious scripts, exploit instructions, malware, or phishing kits. section 5.3
- Subscription packaging and turnkey interfaces can lower the barrier to cybercrime, especially for inexperienced users. section 5.3
- Claims about specialized capabilities may be exaggerated or fraudulent, with some underground tools potentially functioning as scams. section 5.3
