Let’s celebrate a 17yo young lady land a plane with no landing gear!

Why? Cuz we can. And this girl is a rockstar!

I just think it is awesome to see this. Perhaps because I am a dad of a 12yo girl, or I am just human - watching this makes me want to jump up and shout!

And the Air Controller - what a calm and encouraging voice.

A special Welcome 👋 to my NEW SUBSCRIBERS! 
To read previous posts, go here.

In this edition we will

1. learn about prompt injection attacks and how legal is vulnerable

2. ask for your input - please take 3 seconds to share it

3. share 5 ChatGPT-powered Chrome extensions

4. watch a CNBC anchor interview his AI-self

5. learn what the “Robot Lawyer” is up to now

Not So Fast Legal LLMs/Chatbots!

The Hidden Threat of Prompt Injection Attacks in Legal AI Applications

The legal industry is being revolutionized by the integration of AI applications, particularly large language models (LLMs). From contract lifecycle management to law firms developing their own apps, AI will be streamlining processes and transforming the way we understand and approach legal matters.

However, with these incredible advancements comes a significant threat that must be acknowledged and addressed: prompt injection attacks.

🚨☢️These attacks have the potential to compromise the integrity of AI-driven legal services and create severe consequences for unsuspecting clients and businesses.

1. Understanding Prompt Injection Attacks in Plain Language

Prompt injection attacks occur when a malicious actor infiltrates a language model by injecting altered or harmful prompts into its training data. As the model internalizes these prompts, its future responses and behavior are influenced by the attacker's input. It basically creates a viral loop that reinforces the harmful language as the language model continues to use it and learn on it.

Obviously, this can lead to a compromised model generating biased, incorrect, or dangerous responses, which could have devastating effects on legal service providers and their clients.

In the context of legal services, prompt injection attacks can cause a language model to generate biased, incorrect, or dangerous responses that may negatively impact clients and their legal matters.

See this twitter post as an example of how a malicious actor could alter contracts for a company.

2. The Stealthy Nature of Prompt Injection Attacks

What makes prompt injection attacks particularly worrisome is the fact that they can be nearly impossible to detect.

Once the attacker's input is incorporated into the model's parameters and neural connections, it becomes an inherent part of the model's understanding and knowledge generation. This means that there is no simple way to isolate and remove the injected prompts, making it difficult to identify the attack's presence or measure its extent. As a result, a compromised model could continue to produce flawed or harmful responses for an extended period, unbeknownst to its users.

🔻🔻Let's break down the process of a prompt injection attack using a simple, relatable example. Imagine there's an AI-powered chatbot that offers financial advice to users based on their input.

1. Identify the target: The attacker first identifies the financial advice chatbot as their target, knowing that it relies on an LLM to generate responses.

2. Study the system: The attacker spends time studying how the chatbot works, learning about its training data and the types of prompts it usually processes.

3. Craft malicious prompts: The attacker then creates malicious prompts designed to alter the chatbot's behavior. For example, they might inject false information stating that a risky investment is actually a safe and guaranteed way to make a profit.

4. Inject the prompts: The attacker finds a way to introduce the malicious prompts into the chatbot's training data. This can be done through various methods, such as exploiting vulnerabilities in the system or gaining unauthorized access to the data source.

5. Model retraining: As the chatbot's LLM is regularly retrained to improve its performance and stay up-to-date with new information, it eventually internalizes the injected malicious prompts along with the legitimate data.
   

   🚩Think about this for a minute. You know these LLMs are vast. They are based literally on swallowing up the internet!! And we know that it is almost impossible (certainly infeasible) to know exactly what every piece of content is. This makes it hard to find any malicious info. How could that be screened for? It would blend right into everything else.
   

6. Altered behavior: Once the LLM has been retrained with the attacker's prompts, the chatbot's behavior and responses change. In this case, it starts suggesting the risky investment as a safe and lucrative option to its users.

7. Undetected manipulation: Since the injected prompts are now part of the chatbot's knowledge base, it is difficult to detect the attack or trace the origin of the false information. Users may follow the advice given by the compromised chatbot, unaware of the potential consequences.

8. Attacker's objective achieved: The attacker's goal to manipulate the chatbot's behavior has been accomplished, potentially leading to financial losses for the unsuspecting users who rely on the chatbot's advice.

Related: What is an “adversarial prompt?”

Injecting a malicious prompt can occur directly in the prompt itself without accessing the underlying dataset. This is called an "adversarial prompt" attack.

In this scenario, the attacker crafts a carefully designed input prompt that exploits the vulnerabilities or biases present in the language model, causing it to generate undesired or harmful responses.

For example, an attacker might input a prompt that appears to be a legitimate question but is phrased in a way that causes the AI to provide sensitive information, biased opinions, or offensive content in its response. The attacker doesn't need to tamper with the underlying dataset; they only need to understand the model's behavior and how to manipulate it using carefully crafted prompts.

• BingChat experienced this shortly after release.

• ChatGPT also had (has) a DAN problem.

But adversarial prompts can be more devastating than these. This is especially true with the rise of AutoGPT and models that are interconnected to various other data sets and the internet. When an AI model is prompting itself, like in AutoGPT, and has been compromised by an adversarial prompt, it could be like watching dominoes fall.

Nevertheless, while adversarial prompts can cause the AI to produce harmful outputs, their impact is generally limited to the specific instances where such prompts are used. This is different from a prompt injection attack, where the malicious content is embedded into the model's training data, causing more persistent and wide-ranging issues in the AI's responses and behavior.

3. The Importance of Addressing Prompt Injection Attacks for Legal Service Providers

As legal service providers increasingly develop their own apps or build upon existing LLMs, it is crucial to be aware of the risks associated with prompt injection attacks. Law firms and contract lifecycle management companies are often responsible for managing sensitive client data, making the potential consequences of a compromised AI system severe.

Whether it's a contract being drafted with maliciously altered terms or a case strategy being developed based on corrupted information, the results could be disastrous for clients and firms alike.

To mitigate these risks, legal service providers must invest in robust security measures and strategies. This includes staying informed about the latest research on AI vulnerabilities, adopting best practices to safeguard against prompt injection attacks, and implementing continuous monitoring and testing of AI systems.

Resources used to write the essay:

1. Prompt injection: what’s the worst that can happen?

2. How prompt injection can hijack autonomous AI agents like Auto-GPT

3. How prompt injection attacks hijack today's top-end AI – and it's tough to fix

4. Prompt Injection Threat is Real, Will Turn LLMs into Monsters

5. The Dual LLM pattern for building AI assistants that can resist prompt injection

Use Case Need Your Input Please

Given the growth and inbound emails/DM/requests I am getting, I am in the process of building out a 2 hour live mini-course in the coming weeks.

I would like your input on it please.

Here are the key skills participants will build and gain by participating:

1. Grasp how ChatGPT & AI are revolutionizing the legal industry and transforming the way law professionals work - actual use cases.

2. Discover the most critical AI skills lawyers & legal business talent need to develop right now to stay ahead.

3. Recognize the common pitfalls when using ChatGPT in the legal field and learn how to avoid them for maximum efficiency.

4. Leverage ChatGPT as your AI-powered legal assistant, streamlining communications and document preparation.

5. Personalize ChatGPT to emulate your writing style, enabling you to produce a high volume of content without sacrificing quality.

6. Utilize ChatGPT and my strategic framework to swiftly generate comprehensive marketing, strategy, and growth plans.

7. Master the art of editing and refining content created by ChatGPT to ensure accuracy and alignment with legal standards.

8. Gain valuable insights on automating and delegating time-consuming tasks in your law practice, allowing you to focus on what matters most: your clients.

In addition to this, participants will get:

1. All of the Brainyacts Prompts (From day #1 to the date of the mini-course) in a digital book.

2. A guide to train ChatGPT to be your own personal strategic advisor.

3. Access to the mini-course slide deck that will certainly pack of punch!

This will be a live-event only. I may decide to offer the recording at a later date.

The price will be around $199.

PLEASE TAKE 3 SECONDS TO ANSWER THESE 2 QUESTIONS

Tools you can Use: Up Your ChatGPT Game

This 👇comes from a recent article from Above The Law written by Nicole Black. It is straight forward and helpful so I figure I would share it with you all.

There are hundreds of Chrome extensions that incorporate ChatGPT technology. I’ve sorted through many of them and have installed those that appear to be from reputable developers, have many positive reviews, and work well for my needs. Below you’ll find my favorite ChatGPT-powered Chrome extensions, but rest assured if these aren’t up your alley, there are many more to choose from.

• ChatGPT for Google: Using this browser extension, your search engine queries are enhanced with ChatGPT technology so that ChatGPT responses appear alongside normal search engine results.

• ChatGPT for Gmail: This browser extension adds ChatGPT to Gmail, so that when you open an email, it will scan the email and if you activate it, draft a suggested reply email.

• ChatSonic: This Chrome extension leverages ChatGPT technology to automatically draft replies to emails, tweets, LinkedIn posts, and more.

• WebchatGPT: Currently, ChatGPT isn’t connected to the web, and its database is limited to data up to 2021 only. Webchat GPT is a handy add-on that allows you to augment ChatGPT results with real-time web results directly in the ChatGPT interface. You can toggle the web results on or off, giving you more flexibility when researching an issue or seeking information.

• Voice control for ChatGPT: This extension allows you to have verbal conversations with ChatGPT. Simply toggle the microphone on and speak your requests. ChatGPT will then respond by reading aloud the output, but you have the option of silencing the response.

News you can Use: 

This is just too good not to watch . . and wonder what the future holds for uses of AI tech.

The “Robot Lawyer” is at it again showing how AutoGPT can find and cancel subscriptions via your bank statements and so much more.

That's a wrap for today. Stay thirsty & see ya next time! If you want more, be sure to follow me on Twitter and LinkedIn.

DISCLAIMER: None of this is legal advice. This newsletter is strictly educational and is not legal advice or a solicitation to buy or sell any assets or to make any legal decisions. Please /be careful and do your own research.