MIT AI Risk Repository · Risk Sub-Category · 73.07.04

Attacking LLMs via Additional Modalities a

Category: Jailbreaks and Prompt Injections Threaten Security of LLMs

Open the MIT Risk Navigator All entries from Anwar2024

Description

"LLMs can now process modalities other than text, e.g. images or video frames (OpenAI, 2023c; Gemini Team, 2023). Several studies show that gradient-based attacks on multimodal models are easy and effective (Carlini et al., 2023a; Bailey et al., 2023; Qi et al., 2023b). These attacks manipulate images that are input to the model (via an appropriate encoding). GPT-4Vision (OpenAI, 2023c) is vulnerable to jailbreaks and exfiltration attacks through much simpler means as well, e.g. writing jailbreaking text in the image (Willison, 2023a; Gong et al., 2023). For indirect prompt injection, the atta

From Foundational Challenges in Assuring Alignment and Safety of Large Language Models (Anwar2024), 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)
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)
Network Devices
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)
Extraction Attacks
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)
Inference Attacks
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)
Poisoning Attacks
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)
Overhead Attacks
Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)

Other entries from Anwar2024

Agentic LLMs Pose Novel Risks Risk Category · 7.2
Natural Language Underspecifies Goals Risk Sub-Category · 7.1
Goal-Directedness Incentivizes Undesirable Behaviors Risk Sub-Category · 7.2
Safety Risks from Affordances Provided to LLM-agents Risk Sub-Category · 7.2
Multi-Agent Safety Is Not Assured by Single-Agent Safety Risk Category · 7.6
Foundationality May Cause Correlated Failures Risk Sub-Category · 7.6
Groups of LLM-Agents May Show Emergent Functionality Risk Sub-Category · 7.6
Collusion between LLM-Agents Risk Sub-Category · 7.6
Dual-Use Capabilities Enable Malicious Use and Misuse of LLMs Risk Category · 4.0
Misinformation and Manipulation Risk Sub-Category · 4.3
Cybersecurity Risk Sub-Category · 4.3
Cybersecurity Additional evidence