AI vs. AI: Good Artificial Intelligence Still Defeats Evil

Even the most modern security systems do not guarantee unquestionable protection of business data because cybersecurity can be metaphorically defined, in addition to complex IT language, as a race against time and technology. There is a goal – security, but there is no limit; the winner is the one who constantly follows trends in new technologies and creatively utilizes available technological tools in real time, especially artificial intelligence (AI), which can help – when it recognizes and prevents unauthorized access to the business system, or hinder – if misused by those with bad intentions. This is why large language models, large language model (LLM) and open-source programs have come to the forefront, as some of them are not only effective but also accessible and inexpensive. They are developed by large tech companies as well as startups.

Open Models

Mistral AI, operating in Paris, recently introduced its first large language model, Mistral 7B, and immediately announced new ones. It can be easily downloaded from GitHub, and compared to similar models like Llama 2, Mistral 7B offers similar or better capabilities than, for example, models like GPT-4, which are not easy to use as they are mostly available via APIs (from English application programming interface). Mistral 7B is free to download and use, optimal for low latency, text summarization, classification, and text and code completion. Mistral AI decided to release Mistral 7B under the Apache 2.0 license, which has no restrictions on use or reproduction.

‘Working with open models is the best way for both suppliers and users to build a sustainable business around AI solutions. Open models can be tailored to address many new fundamental business problems across all industrial verticals, unmatched by black box models,’ stated Mistral AI in a blog post published on the occasion of the Mistral 7B launch.

The blog also states that in the future, they will develop many different specialized models, each tailored to specific tasks, compressed as much as possible, and connected to specific modalities, particularly targeting business clients and their needs related to research and development, customer care, and marketing, as well as providing them with tools to create new products using artificial intelligence.

Additionally, the transparency of Mistral 7B due to decentralization is highlighted – this model alone, they claim, can generate a response to any query without moderation. Compared to models like GPT and Llama, whose mechanisms determine when to respond, Mistral 7B is free in choice, or decentralized, but the problem is that this can be exploited by bad actors. Nevertheless, the potential danger of misuse does not diminish its advantage as its overall benefit is much greater because it can make artificial intelligence accessible to everyone.

Fake Records

Mistral 7B is available under the Apache 2.0 license, meaning there are no barriers to its use: it can be used personally, in large companies, or in public administration. It only requires ensuring an appropriate system for its operation, being cautious with data, educating employees, and being prepared for the mischief of artificial characters like deepfakes and chatbots, and mitigating the damage that may arise from using artificial intelligence.

—

—

– The most important, yet less obvious vulnerability in using artificial intelligence is data security. Advanced AI models imply processing large amounts of data, sometimes personal, raising questions about the security of processing and storing that data in the cloud. End users will also, without much hesitation, copy confidential information or part of the source code into a chatbot interface to get help with translation or solving a business challenge, thus opening the possibility of sensitive data leaking outside the business environment – explains Filip Kiseljak, an information security expert at KING ICT. According to him, security issues can also arise from deepfake and chatbot technologies, which are advancing day by day and have become easily accessible in recent years, requiring no special expertise for their use.

– It is becoming increasingly difficult for human eyes and ears to distinguish deepfake video and audio clips from authentic ones, and AI technology has begun to be used for the detection of such works. Deepfake technology opens up possibilities for malicious third parties for fraud, identity theft, spreading misinformation, and causing reputational damage by creating fake audio or video recordings. There are examples where deepfake technologies have been used to create false statements from various politicians, successfully spreading misinformation and causing instability in global political spheres – lists Kiseljak.

Questionable Privacy Protection

He adds that the use of chatbot technology is primarily associated with advantages such as fast analytics and problem-solving, time savings, high availability…, but some of the problems it brings are inaccurate and biased responses, questionable privacy protection, security vulnerabilities, and high implementation and maintenance costs. According to Kiseljak, a large number of users unreservedly accept the responses they receive in conversations with chatbots, which can lead to incorrectly applied solutions in business systems and may result in errors or even business interruptions.

– AI is not only used in business for constructive actions. Globally, automated attacks and the development of advanced malware using artificial intelligence (AI) are on the rise, and security software manufacturers are also reporting an increase in malware-free attacks (malware free) from approximately 60 to 70 percent of all detected attacks on business systems. Customized AI models can also be used for automated vulnerability discovery and planning attacks on targeted business environments – notes Kiseljak.

He emphasizes that it is known that attackers are always one step ahead of defenses, but it remains to be seen how much the use of artificial intelligence (AI) and advanced machine learning (ML) will help or hinder attackers in their efforts.

– A large number of established security software manufacturers with whom we closely collaborate at KING ICT have incorporated some form of AI-driven analysis into their products, enabling security analysts to more easily triage and detect real threats, as well as automated responses in the case of high-risk threats, providing organizations and companies with a higher level of protection, better insight into security status, faster adaptation to threats, and increased efficiency of the security operations center – explains Kiseljak.

Three Main Problems

According to Alen Delić, Vice President of the Croatian Association of Security Managers, the space for technology development is extremely broad, and only a part of that development and use today falls under what is collectively called artificial intelligence.

– First of all, we should know for what purposes we need or want to use different tools, including those that do not use AI as we perceive it today. The use of such tools carries various risks and problems. Some of them are legal in nature, such as data privacy, copyright issues, and so on. Secondly, from the perspective of risks associated with third-party use, we forget about the area of testing and verifying tools, which puts us in a position to use tools we do not know how they work and what risks they carry. And lastly, perhaps something we forget, is the area of education and awareness of users who use such tools. The use of tools should also entail changes in the way of educating and training users so that they know what the capabilities of such tools are and what risks they entail. Or, as ChatGPT itself would say when we ask it: ‘Introducing artificial intelligence into business can bring numerous advantages, but it also carries serious security challenges. Organizations must be proactive in identifying these risks and implementing appropriate security measures to protect their data, resources, and reputation.’ – emphasizes Delić.

Root of Vulnerability

He also emphasizes that most employees today actually only use tools that help them speed up some automated tasks, and vulnerabilities in using artificial intelligence in business are mostly related to unreliable data they may use, entering confidential or personal data into systems they do not know how they function and for what the data they enter is used, as well as certain biases of models that users do not verify.

– Deepfakes and chatbots are automated tools that can be used for fraud and manipulation of information, which can consequently affect the reputation of individuals and organizations or states – emphasizes Delić, adding that they can be used to spread misinformation, collect sensitive information through social engineering, as well as for other attacks that involve the need for rapid information exchange.