In May 2024, Microsoft revealed that it had developed a generative AI model completely isolated from the internet, noting that US intelligence agencies could safely use this powerful technology to analyze highly classified information they work with. This indicates that the revolution launched by OpenAI in the field of large language models through the release of ChatGPT is reaching intelligence agencies, which have since been working to employ the great capabilities of language models and AI techniques in general for intelligence tasks, in a way that does not violate privacy and security standards followed in intelligence agencies.
Microsoft’s Model
Microsoft announced the new product at the National Competitiveness AI Expo held in May 2024 in Washington, D.C., USA. The key features and implications of Microsoft’s new intelligence model can be summarized as follows:
1- US intelligence seeking to benefit from AI: US intelligence has clearly expressed its desire to benefit from new AI technologies in intelligence tasks. Intelligence officials previously confirmed in February 2023 that there is a lot of work underway within the CIA to ensure it becomes an AI-driven organization, in addition to the parallel aspect of studying and understanding adversaries’ use of AI and machine learning capabilities. The CIA has also created a joint platform to activate and eventually expand AI applications, and create new resources and opportunities to increase its employees’ familiarity with smart and automated technologies.
Lakshmi Raman, Director of Artificial Intelligence at the CIA, noted in her assessment of AI programs and the potential for US intelligence to benefit from them in September 2023 that this powerful technology also has potential positives for the United States, emphasizing that the roles of these applications are not limited only to potential security threats, and that they have capabilities that can be leveraged. She pointed out that the key advantage of AI is that it is “widely available and works at a speed and ease that was not possible before.”
2- Developing the first isolated and offline model: While the CIA began using a generative AI tool for non-classified purposes in 2023, the need to isolate the platform from the internet to improve its security against cyber threats and ensure that sensitive national security information does not leak was a driver towards developing a new model.
Microsoft’s generative AI model for US intelligence aims to overcome the security problems that stem from large language models (LLMs) being connected to the internet, which is usually used as a resource for training those models. William Chappell, Chief Technology Officer for Strategic Missions and Technologies at Microsoft, has already announced that this is the first time the company has developed an isolated and offline version, which is one of the main differences between this model and other AI models that are vulnerable to hacking and data leaks.
Although the model also technically relies on the cloud like other smart models and applications, what is different in this case is that the model developed by Microsoft is connected to a private network that can only be accessed by the US government and intelligence agencies concerned with it. It is closer to a “data warehouse” or “closed storage network,” which Chappell described as a static cloud environment based on “air-gap” technology and isolated from the internet. Thus, the model cannot learn from data and files shared by other users or information derived from the web, but rather relies only on its own network, and does not share its own information and data outside that network.
3- Relying on “GPT-4” in building the new model: Microsoft’s model is considered a means of processing and analyzing huge amounts of classified information without compromising confidentiality. Bloomberg had already revealed in September 2023 that the CIA was building its own internal ChatGPT-style tool to examine public data. According to what has been announced, Microsoft’s intelligence model is built on OpenAI’s GPT-4 large language model.
As is known, ChatGPT is a “chatbot” program developed by OpenAI, which spread rapidly since its launch in late 2022. It is part of generative AI that includes large language models that can create audio, symbols, images, texts, videos and other media. GPT-4 is the fourth version of the model, launched in March 2023.
4- Maintaining secrecy on details of the US project: Despite all this, the information available about the new generative AI model revealed by Microsoft for US intelligence is still limited. The company was keen to reveal its new model to promote it as a technological success and precedent for the company, without revealing more details about the model, especially regarding technical and operational aspects. This is related to the model’s customization for US intelligence, and the great secrecy that requires, and the fear of other intelligence agencies benefiting from some technical details in efforts to develop their competing models, or understanding the mechanism of the US intelligence model accurately, or working to undermine and harm it.
5- Starting to test the model within the intelligence community: The new American large language model underwent an 18-month development process by the development team responsible for the project at Microsoft, for the benefit of US intelligence. Despite the company developing the intelligence model announcing it and completing work on its development, it is still in the experimental work phase, as the model is currently being tested by various agencies of the US intelligence community. This means that the model is still in the development stage and has not been fully completed or its work generalized. The new AI platform of Microsoft for US intelligence agencies is going through the testing and approval phase before it enters regular use, and before the US intelligence community uses the new model on a wider scale. The American company will first make some modifications and address some problems revealed by the ongoing testing process, before generalizing its use fully by US intelligence.
Determinants of Development
The success of US intelligence in developing a language model based on AI technologies for use in some of its tasks opens the door for various other intelligence agencies in countries around the world to look towards benefiting from these emerging technologies and models. Along with the great international competition to develop AI technologies in security and military fields, there is competition between intelligence agencies to develop and benefit from large language models after the great leap achieved by the ChatGPT model, while maintaining known intelligence work standards, foremost of which is privacy. Among the most prominent determinants and considerations governing the orientation of various intelligence agencies in the world towards developing and using generative language models in their work are the following:
1- International race for generative AI: The ability of generative AI to analyze huge amounts of data and recognize patterns to provide users with scientific evidence, analysis and actionable insights has made such tools highly sought after by intelligence agencies, and has driven competition between major intelligence agencies in the world to benefit from technologies and models based on generative AI as an emerging field that contains many opportunities that intersect with the work of intelligence agencies, especially with regard to data processing.
Sheetal Patel, Assistant Director of the CIA for the Center for Transnational and Technology Missions, said at a security conference at Vanderbilt University in April 2024 that there is a race to obtain, develop, and benefit from generative AI in processing intelligence data, and that intelligence agencies around the world are already competing to be the first to harness generative AI in their work. She revealed the United States’ ambition to be the first country to succeed in using generative AI in this field, considering that precedence in this field will create differences in work, development and analytical capabilities between US intelligence and other intelligence agencies in the world.
2- Intelligence interest in developing large language models: US intelligence has clearly expressed its desire to benefit from new generative AI models, especially after OpenAI launched the ChatGPT chatbot, which revealed huge potential and opportunities to benefit from generative AI models and their ability to understand, analyze and absorb huge information, as well as machine learning capabilities.
The CIA’s AI team has also intensified its interest in how ChatGPT and similar capabilities can provide more advanced assistance to agency employees while carrying out important intelligence operations. Lakshmi Raman, Director of Artificial Intelligence at the CIA, previously confirmed in September 2023 that the agency is working on developing an internal chatbot to assist with research and writing.
3- Analyzing huge amounts of intelligence data: Intelligence agencies seek to benefit from the technical leap created by AI technology, especially large language models (LLMs) like ChatGPT. The new smart models that will be developed and used by major intelligence agencies in the world will seek to help their agents and analysts by reviewing huge amounts of data and identifying key trends, a process that intelligence officials confirm would not be possible to apply to huge amounts of data without using AI techniques and models.
While Microsoft’s model is a means of processing and analyzing huge amounts of classified information without compromising confidentiality, other intelligence agencies are likely to use similar models to analyze huge intelligence data in search of evidence, taking advantage of huge intelligence databases, whether within one agency, or within the intelligence community in the state in general, or even intelligence databases that are built through multilateral intelligence information exchange mechanisms.
4- Expanding the development of closed models for security considerations: Most AI models rely on cloud computing for machine learning and inferring patterns from data. But US intelligence has moved towards developing models isolated from the global internet, to improve their security against cyber threats and ensure that sensitive national security information does not leak into a model that can be accessed publicly. The same is expected to apply to the various AI models that intelligence agencies in different countries of the world seek to develop and use, within the framework of intelligence agencies’ interest in benefiting from the great capabilities of AI in understanding and analyzing the huge amounts of classified information possessed by these agencies, while at the same time ensuring data security and privacy.
Microsoft considered that developing generative AI models separately from open internet networks and cloud computing enables intelligence agencies to safely use smart technology to analyze highly classified information they work on. It is worth noting that one of the main tasks undertaken by AI teams in intelligence agencies is understanding the protection barriers required when applying emerging smart technologies to intelligence community operations, to benefit from the great advantages provided by these technologies and models, without violating privacy and security standards followed in intelligence agencies.
While cloud-based models involve transferring data to external servers for processing, which raises security concerns, storing and processing sensitive information on internal devices and networks can significantly reduce the risks of unauthorized access and potential data breaches. Therefore, Samsung, for example, restricted the use of ChatGPT on the company’s devices to a minimum, and worked on building its own chatbot to reduce these risks, after three of the company’s employees caused sensitive data to be leaked to ChatGPT. Other institutions such as Amazon, Verizon and JP Morgan have also restricted the use of ChatGPT in their daily operations due to security concerns, as the model collects a lot of data from users, including all conversations the user has with the platform, and it is trained using the data it receives, including confidential company or intelligence agency data.
This is a matter of high sensitivity and danger in intelligence work, as intelligence agency data includes satellite images, videos, and highly accurate classified information, and if these programs obtain this data and analyze it or show its results to general users, or to other intelligence agencies, this will cause big problems for the intelligence agency, capable of thwarting its work and threatening the national security of the state.
5- Granting specific permissions to access model data: As mentioned earlier, the AI model developed by Microsoft for US intelligence is on a private network that can only be accessed by the US government and the US intelligence agencies concerned with it. The Chief Technology Officer for Strategic Missions and Technologies at Microsoft pointed out that about 10,000 people will theoretically be able to access the model developed by his company.
Due to the nature of their tasks and the confidentiality of their information, AI models that major technology companies will develop for the intelligence agency will be closed models, with specifying who has the authority to enter them, as they will be limited to employees of intelligence and public security agencies in the state. They can also be closed to employees of some departments; as determining access permissions is related to the factor of privacy and security, and is also related to ensuring that important intelligence information that will be stored on the model’s network is not leaked.
6- Dispensing with machine learning capabilities from data: While many look forward to benefiting from the open horizons of AI applications, foremost of which is the ability to machine learn and benefit from big data and open sources; The privacy and reliability issues that accompany these broad features are issues that intelligence agencies cannot afford to bear in view of the pivotal importance of the accuracy of analyzes and data confidentiality in the work of these agencies. It was noteworthy that William Chappell, in his statements to Bloomberg, emphasized that the new AI platform developed by Microsoft is designed so that it can read files, but does not learn from them in a way that may affect its production or from the wider internet. This indicates that one of the main differences between special intelligence models and their open general counterparts is the technical orientation during the development process that AI models developed for the benefit of intelligence agencies are not able to acquire machine learning capabilities, and do not learn from questions asked by users, in a way that later contributes to revealing intelligence information. They will be simple and direct models that answer questions, and closer to models that write codes. Intelligence agencies are moving towards not giving these platforms the ability to learn from the data they are provided with, and for large language models dedicated to intelligence work to process each case separately.
7- Relying on available commercial tools: As the CIA did in the early stages of building its intelligence model, by researching the best models available, until it decided to rely on GPT-4, it is likely that other intelligence agencies in countries of the world will do something similar, taking into account that the choices will be fewer in terms of number compared to the huge American market in the field of AI, especially since the factor of reliability and relationships with companies or research centers responsible for these technologies is extremely important.
The CIA has worked collaboratively with entities that play a pivotal role in the field of generative AI, including federally funded research and development centers, major companies and other partners in academic and industrial circles. It collaborated with all of these to evaluate a large number of available commercial tools and products, to introduce them into the intelligence sector with some necessary development and adaptation processes.
To reduce financial cost and time periods necessary to develop intelligence programs and tools, intelligence agencies will work to integrate externally built tools developed by other companies and link them to secret and protected computer systems used in intelligence agencies and concerned with national security affairs. This may also be the only option for intelligence agencies that do not have or their direct partners within the state have great technical expertise in this field.
8- Companies seeking to gain the trust of intelligence agencies: On the other hand, the major technology companies themselves, specialized in AI technologies and large language models, will try to take advantage of intelligence agencies’ tendency to rely on these models, to obtain potential investment gains by working with intelligence agencies in developing and operating these models, with the requirements that this entails, foremost of which is enhancing the confidence of intelligence agencies in these companies.
These companies, including Microsoft, will seek to present themselves as companies assisting and cooperating with the government military and intelligence sector in providing these entities with generative AI services, and developing large language models for these agencies to use in their work, while reassuring them by providing greater advantages in terms of privacy and security factors, as they are among the biggest challenges facing companies working in this field, specifically general generative AI programs connected to the internet, such as ChatGPT or Gemini.
9- Concern about the powers of companies developing models: It should be noted that the approval and accreditation process for new contractors and technical partners is often long and arduous for security and bureaucratic considerations. Many intelligence officials believe it should be updated and simplified, in a way that does not violate the required degrees of security and reliability as well. But this matter is surrounded by a bigger debate not only about the reliability of relying on these companies in purchasing and using the technologies they have developed, but also about what relates to the powers of these companies to access their tools and associated clouds.
The powers of companies developing AI models used in the intelligence field, and the ability of these companies to access data repositories will remain a matter of dispute. No matter how much confidence increases in companies developing AI models, it is expected that parties will argue that despite the military and intelligence sectors benefiting from the technological boom achieved by the civilian AI sector, this does not negate the necessity for military and intelligence institutions to work on developing the smart technologies and tools they need themselves through technological arms that they establish and control, which adds more security and reliability to these tools and the data they are provided with.
Others believe that there is no need for this expansion with its great cost and complex tasks, and that “civil-intelligence” or “civil-military” partnership mechanisms serve the purpose, with some development in these tools, and limiting the powers of developing companies after operation to a minimum. They believe that the mechanism of partnership with civil entities is an established mechanism that has controls governing it, and has proven its effectiveness in various fields, including military industries, and that it reduces many burdens on military and intelligence institutions, and allows benefiting from the open horizons of civilian and dual-use technological innovations.
Speed of Achievement
In conclusion, while analyzing huge intelligence data is a large and complex task, and may take a long time and involve a large number of people to obtain the required results; Generative AI technologies and large language models will not only reduce the burden on employees and analysts of intelligence agencies, but they help in understanding and analyzing data and identifying required answers and evidence in a very short time, which is consistent with the importance of the time factor in intelligence work.
