Snowden’s revelations have made it possible to objectively discuss the role of espionage and wiretapping. NASA not only engages in space exploration, but also appears to be responsible for eavesdropping on internet and mobile phone information worldwide. It has also become clear that the United States is also stealing information from its allies. However, the current challenge is not just about hacking the Internet; it is about reading large amounts of data and analyzing it. The 9/11 terrorist attacks are a good example. Wiretapping was successful during the 9/11 attacks. However, the volume of information was so large that it was difficult to properly analyze.
The current challenge in wiretapping is reading large amounts of encrypted data and analyzing it. Encryption is designed to be unbreakable, yet it is increasingly being broken. There’s a reason for this. Current mainstream encryption exploits the time it takes computers to solve prime factorization problems. The idea that a problem is secure unless there is a fast solution is called “computational security.” As more computationally powerful tools become available, the scope of code and cryptanalysis methods expands and the speed of decryption increases. Computational security is destined to be broken as computational technology advances. Computational technology is rapidly advancing with the use of supercomputers and other technologies. Another challenge in wiretapping is language. In Europe, the United States, and Japan, there are few Arabic experts. As a result, counterterrorism measures are often delayed. However, Israel has many Arabic experts, which ensures Israel’s security.
A tool has emerged to solve the challenges of wiretapping and analysis, as well as language issues. It is generative AI. Until now, language models have generally been considered weak at logical and mathematical tasks. However, somehow, when the model size is increased, their capabilities blossom at a certain point, and they suddenly become able to do things that were previously impossible. They become fluent and accurate in logical and mathematical tasks. To achieve good performance, language models can now be overcome simply by training them on large datasets for long periods of time. This has the potential to overcome the barrier of “computational security.” However, when it comes to making language models huge, both the preparation of the dataset and the training are expensive. Funding this is likely to be a challenge facing modern wiretapping.
