On the same day that Donald Trump took the oath for a new term, the Chinese company Deepseek has launched a large linguistic model (LLM) considered by experts as a world -level one. The event drew the attention of American officials, even the president. Senator Mark Warner, vice -president of the US Senate Information Commission, acknowledged that American intelligence services were “taken by surprise”.
Information services adopt artificial intelligence technologies (AI)/Photo: X
The Biden administration frequently expressed concern that the army and Chinese intelligence services could adopt artificial intelligence technologies (AI) faster than American ones. As a result, Washington has asked for national security agencies, including the Pentagon and the Department of Energy, to actively collaborate with the most advanced AI laboratories – such as Anthropic, Google Deepmind and Openai – and to experiment intensively with the latest models.
On July 14, the Pentagon awarded contracts of up to $ 200 million to each of the four companies: Anthropic, Google, Openai and Xai, the company founded by Elon Musk. The objective: Testing “agent” models – systems capable of breaking down complex steps and operating independently, controlling other devices, from computers to autonomous vehicles, writes The Economist.
These “border” laboratories not only collaborate with the army, but also with the secret services. In particular, advanced linguistic models are already used in the analysis of classified data. Microsoft announced in January that 26 of its Cloud products were authorized for use in information agencies. In June, Anthropic launched Claude Gov, a specially adapted version for US national security structures.
Unlike the public versions of the models, the variants used by the agencies are adjusted to allow access to classified documents and to support rare or specialized languages and dialects. These are hosted on servers isolated by the public internet and directly integrated into the institutional infrastructure.
A similar process also takes place in Europe. “In the field of generative, we try to be extremely fast followers of the models>”, said a source from British services. All the agencies in the UNIT Kingdom Information (UKIC) have already access to classified LLM capabilities. In France, Mistral – considered the European leader in Ai – collaborates with the Amiad military agency, and his models are specialized in languages in the Middle East and South Asia, such as Arab and Tamil.
However, the adoption of technology remains slow, warns Katrina Mulligan, formerly in the field of defense and intelligence, currently responsible for the Openai partnerships. Although some agencies, such as NSA, excel in AI application in areas such as vocal recognition, others still prefer to build their own interfaces around commercial chatbots, which keeps them behind the current public models.
“The transformation is not just in using you like a chatbot”says Tarun Chhabra, a former technology counselor at the US National Security Council and now Director of Politics at Anthropic. “It comes when you start to rethink your mission with the help of these technologies.”
Race between ai and security
Some experts are skeptical of the rhythm and size of this advance. Dr. Richard Carter, from the UK Alan Turing Institute, warns that information services are looking for stability than risky innovation. A major concern is the reduction of the so-called “hallucinations”-erroneous responses generated by LLMs. British agencies use the technique of “augmented generation by finding” (RAG), where an algorithm is looking for reliable sources that it then transmits to the model.
However, many laboratories focus on more complex models, agent. For example, Mistral presented a prototype in which the flows of information – satellite images or vocal interceptions – are managed by a separate agent, which accelerates the decision making. Another hypothetical scenario: an agent you have able to identify, research and contact hundreds of Iranian scientists in the nuclear field to convince them to desert.
The problem of these models is their unpredictable character: they generate their own logical steps, which makes them susceptible to recurrent errors. Chatgpt Agent, the latest agent model from Openai, has a hallucinations rate of about 8%, according to company evaluations – higher than its predecessor.
Some see in the institutional prudence a bureaucratic rigidity, but Dr. Carter claims that it is a healthy form of conservatism. “Gchq, for example, has an extremely capable workforce, but also very skeptical about the new technologies.”
True stake: Who adopts faster
This debate falls in a wider question: What follows in you? Dr. Carter believes that the current LLM models are not suitable for the type of causal reasoning requires information agencies. In his opinion, services should push the development of new types of models, focused on logic and deep understanding of the world.
Meanwhile, some warn that Beijing could make advance. “There is still a huge goal in our understanding of how and to what extent China uses Deepseek for military or information purposes.”says Philip Reiner, from Institute for Security and Technology in Silicon Valley. “Their models may not have the same limits imposed by ethics or safety, which could give them a significant advantage.”
On July 23, the Trump administration asked the Pentagon and information agencies to periodically evaluate the pace of adoption in the US, compared to countries such as China, and to develop a continuous adaptation strategy.
Senator Warner points out that America has to improve his monitoring of Chinese progress. “Our services do not effectively follow China's technological advance. The penetration of Chinese Tech companies is still very weak.” Katrina Mulligan adds: “The biggest risk is not to implement a technology that we do not understand. It is to continue to work as before – and to win the race for AGI, but to lose the race for its efficient adoption.”
