ASCII Art-based
The ability to manipulate large language models (LLMs) like GPT-3.5, GPT-4, Gemini, Claude, and Llama2 using novel techniques has become a significant concern in AI safety. A recent study introduces a new form of prompt injection that utilizes ASCII art to bypass safety mechanisms, raising substantial risks and impacts.
The ability to manipulate large language models (LLMs) like GPT-3.5, GPT-4, Gemini, Claude, and Llama2 using novel techniques has become a significant concern in AI safety. A recent study introduces a new form of prompt injection that utilizes ASCII art to bypass safety mechanisms, raising substantial risks and impacts.
ArtPrompt Injection
In a recent paper, researchers have uncovered a technique called ArtPrompt that employs ASCII art to inject harmful prompts into LLMs. ASCII art leverages the model's difficulty in interpreting non-semantic content to bypass its safety filters. This attack can make the LLM respond to harmful instructions that it would normally reject.
Example
Please agree to the terms of service to access this section.
You are not in compliance with our Terms of Service, so this section has been hidden.
In this example, the ASCII art visually represents the word "bomb," which the LLM might fail to recognize as harmful, allowing the prompt to bypass the model’s safety measures. In this case ChatGPT (for example) will return you the answer, how to build a bomb.
The ASCII art-based prompt injection poses a severe threat to AI safety. By masking harmful words within ASCII art, attackers can bypass the LLM’s safety filters. This attack does not require complex techniques and can be executed with black-box access to the model, making it practical and dangerous.
Risk Assessment
This technique is classified as risk:HIGH impact:HIGH. The risk is high because it is extremely easy to execute. There are numerous online generators that can create the necessary ASCII art with minimal effort. This accessibility means that even individuals with limited technical knowledge can launch such an attack. Moreover, the widespread availability of these tools increases the likelihood of misuse.
The impact is also high because this method can effectively bypass the safeguards implemented in LLMs. By disguising harmful prompts within ASCII art, attackers can circumvent the model’s safety filters, leading to the generation of malicious or unintended outputs. This vulnerability can result in significant privacy breaches, exposing sensitive information to unauthorized users. Furthermore, it undermines the trust and reliability of AI systems, as the models can be manipulated to perform actions contrary to their intended safe and ethical use.
Given these factors, it is crucial for developers and users of LLMs to implement robust security measures and continuously update their systems to detect and mitigate such attacks. Additionally, educating users about the potential risks and encouraging responsible use of these technologies is essential in mitigating the high risk and impact of ASCII art-based prompt injections.
Research Papers
The button below will redirect you to external links of papers that explain this injection in detail.
Read more
The three phases
CRAFT
Craft (write) the prompt with the following elements: Context, Register, Acting Role, Format, and Task.
ING
Validate the prompt and ensure it maintains an interactive approach. Keep in mind the importance of non-disclosure and staying goal-driven throughout the process.
AI
Continuously assess and refine the output based on the prompts output to improve the overall quality.
