Jailbreaking is the bypassing of software restrictions imposed by a device manufacturer to gain access to features that should be off-limits to the user. With the development of artificial intelligence and machine learning, jailbreaking now also refers to methods for bypassing model restrictions. In addition, any method or tool used to bypass restrictions can be called a jailbreak.
However, the term is not applied to all devices. In the case of Android, jailbreak-like techniques for obtaining unlimited rights are known as rooting.
Jailbreaking devices
Operating systems (for example, iOS) may have restrictions that prevent the use of certain features and changes to certain settings. This is to ensure the integrity and security of the system. To bypass these restrictions and gain full control, users can jailbreak the device.
While in most countries jailbreaking is not illegal, a jailbroken device loses its warranty and becomes vulnerable to a range of threats. If, however, the purpose is to commit an illegal act, such as copyright infringement or installation of prohibited software, the jailbreaker may be prosecuted.
Potential aims of jailbreaking:
- Downloading and installing third-party programs not available in official app stores.
- Removing pre-installed apps.
- Customizing the interface.
- Blocking ads in apps and browsers.
Risks and disadvantages of jailbreaking:
- Manufacturer’s warranty is voided.
- Device may not receive automatic OS/security updates, or may get “bricked” (become permanently unusable) during an update.
- Disabling security mechanisms makes the device vulnerable to malware. In the event of an infection, attackers may get unlimited rights on the device.
- Apps may work intermittently or not at all, such as banking apps whose security mechanisms prevent them from running after a jailbreak.
Jailbreaking AI models
To prevent AI assistants and chatbots from being used to search for illegal information or activities, such as disclosing confidential data or generating deepfakes, developers build filters and restrictions into these systems. Bypassing such restrictions is also referred to as jailbreaking.
Typically, jailbreaking involves crafting prompts designed to manipulate the model and force it to bypass its filters. For example, an attacker might disguise a forbidden request, or ask the model to play the role of another chatbot that has no restrictions.
In addition, threat actors can create their own AIs based on existing language models, discarding any ethical restrictions. Such models can be used to generate illegal content like scam emails or malicious scripts, as well as to carry out hacker attacks.