MIT AI Risk Repository · Risk Sub-Category · 62.19.01

Jailbreak of a model to subvert intended behavior

Category: Attacks on GPAIs/GPAI Failure Modes

Open the MIT Risk Navigator All entries from Gipiškis2024

Description

"A jailbreak is a type of adversarial input to the model (during deployment) re- sulting in model behavior deviating from intended use. Jailbreaks may be gen- erated automatically in a “white box” setting, where access to internal training parameters is required for creation and optimization of the attack [238]. Other attacks may be “black box” - without access to model internals. In text based generative models, jailbreaks may sometimes be human-readable, with the use of reasoning or role-play to “convince” the model to bypass its safety mechanisms [231]."

From Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024), as extracted by the MIT AI Risk Repository (CC BY 4.0).

Classification

Domain: 2. Privacy & Security
Subdomain: 2.2 AI system security vulnerabilities and attacks
Causal entity: Human
Intent: Intentional
Timing: Post-deployment

Subdomain definition: Vulnerabilities in AI systems, software development toolchains, and hardware that can be exploited, resulting in unauthorized access, data and privacy breaches, or system manipulation causing unsafe outputs or behavior.

Real-world incidents in this subdomain

Browse all incidents in this subdomain

How other frameworks describe this risk

Privacy loss
A Collaborative, Human-Centred Taxonomy of AI, Algorithmic, and Automation Harms (Abercrombie2024)
Exploiting Limited Generalization of Safety Finetuning
Foundational Challenges in Assuring Alignment and Safety of Large Language Models (Anwar2024)
Jailbreaks and Prompt Injections Threaten Security of LLMs
Foundational Challenges in Assuring Alignment and Safety of Large Language Models (Anwar2024)
“Model Psychology” Attacks
Foundational Challenges in Assuring Alignment and Safety of Large Language Models (Anwar2024)
Attacking LLMs via Additional Modalities a
Foundational Challenges in Assuring Alignment and Safety of Large Language Models (Anwar2024)
Vulnerability to Poisoning and Backdoors
Foundational Challenges in Assuring Alignment and Safety of Large Language Models (Anwar2024)
Model Attacks
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)
Hardware Vulnerabilities
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)

Other entries from Gipiškis2024

Dimension - Intent Risk Category
Intentional Risk Sub-Category
Unintentional Risk Sub-Category
Partially intentional Risk Sub-Category
Dimension - Entity Risk Category
Human Risk Sub-Category
AI Risk Sub-Category
Combination of humans and AI Risk Sub-Category
Dimension - Failure dynamics Risk Category
Isolated (non-normal) failures Risk Sub-Category
System (normal) failures Risk Sub-Category
System (normal) failures Additional evidence