Jailbreak
An attack technique that bypasses the safety alignment mechanisms of large models through carefully constructed prompts, causing them to output content that should otherwise be refused. Along with prompt injection, it is one of the core vulnerability categories in the field of AI safety.
Core Facts
Timeline (last 90 days)
Console jailbreaking generally refers to bypassing the software protection mechanisms set by hardware manufacturers, enabling users to run code that has not been officially signed or authorized on the device.
Console jailbreaking carries a dual significance: it both advances security research and user autonomy (right to repair, homebrew development) and may also be used for game piracy and DRM circumvention.
Console jailbreaks are typically tied to specific firmware versions, and once a user upgrades the system or connects online, the exploit may immediately cease to work.
Attempts to circumvent a model's safety constraints through specific prompt constructions or system-level interventions are referred to as jailbreaking in the field of AI safety research.
Pure prompt injection jailbreak methods are becoming increasingly difficult to execute as vendors continue iterating their systems.
Even if model-level guardrails are bypassed, the model's capability ceiling still exists; jailbreaking more often relaxes the model's tendency to refuse rather than genuinely granting it the ability to complete illegal tasks.
Common jailbreaking techniques include role-play inducement, gradual breakthroughs through multi-turn conversations, encoding-level semantic obfuscation (such as Base64 encoding and Unicode substitution), and adversarial suffix techniques
Common jailbreak techniques include role-play induction, gradual multi-turn conversation breakthroughs, encoding-level semantic obfuscation, and adversarial suffix techniques.
Common forms of prompt injection attacks include direct injection, indirect injection, and jailbreak attacks.
Common prompt jailbreak techniques include role-playing attacks (such as the DAN attack), request decomposition, semantic camouflage, and multilingual switching.
1 more timeline events
All Facts (12)
Model jailbreaking refers to bypassing a large model's safety alignment mechanisms through carefully crafted prompts; common techniques include role-playing attacks, multi-turn conversation induction, and encoding obfuscation.
65%UnverifiedCommon jailbreak techniques include role-play induction, gradual multi-turn conversation breakthroughs, encoding-level semantic obfuscation, and adversarial suffix techniques.
60%UnverifiedThe disclosure of system prompts may make jailbreak attacks easier, as attackers can precisely understand the specific wording and logical structure of the safety rules.
60%UnverifiedConsole jailbreaking carries a dual significance: it both advances security research and user autonomy (right to repair, homebrew development) and may also be used for game piracy and DRM circumvention.
50%UnverifiedConsole jailbreaking generally refers to bypassing the software protection mechanisms set by hardware manufacturers, enabling users to run code that has not been officially signed or authorized on the device.
50%UnverifiedConsole jailbreaks are typically tied to specific firmware versions, and once a user upgrades the system or connects online, the exploit may immediately cease to work.
50%UnverifiedAttempts to circumvent a model's safety constraints through specific prompt constructions or system-level interventions are referred to as jailbreaking in the field of AI safety research.
50%UnverifiedPure prompt injection jailbreak methods are becoming increasingly difficult to execute as vendors continue iterating their systems.
50%UnverifiedEven if model-level guardrails are bypassed, the model's capability ceiling still exists; jailbreaking more often relaxes the model's tendency to refuse rather than genuinely granting it the ability to complete illegal tasks.
50%UnverifiedCommon jailbreaking techniques include role-play inducement, gradual breakthroughs through multi-turn conversations, encoding-level semantic obfuscation (such as Base64 encoding and Unicode substitution), and adversarial suffix techniques
50%UnverifiedCommon forms of prompt injection attacks include direct injection, indirect injection, and jailbreak attacks.
50%UnverifiedCommon prompt jailbreak techniques include role-playing attacks (such as the DAN attack), request decomposition, semantic camouflage, and multilingual switching.
50%