As a computer engineer, I can say with some confidence that defense systems are among the most robust and carefully protected software anywhere. From the outside, with layer upon layer of encryption, isolated networks and tightly controlled hardware, they look close to impenetrable. The real question is whether they are, and whether they will stay that way once artificial intelligence enters the picture.
Start with the basics: passwords and keys
Technically, a password is just a combination of letters, digits and symbols. Suppose you choose a 3-character password from a set of 17 characters: A, B, C, D, E, 0, 1, 2, 3, 4, 5, *, ?, !, #, % and &. That gives 17 × 17 × 17 = 4,913 possibilities, which a computer can try in a fraction of a second. The example is tiny, but it shows the principle: the strength of a secret comes from its length and from how many different characters it can contain.
In serious systems those numbers explode. A 256-bit AES key has 2256 possible values, a number with 78 digits. Even if every computer on Earth did nothing but guess keys, they would not get through a meaningful share of them before the sun burns out. That is why nobody attacks strong modern encryption by guessing.
From Caesar to AES
Encryption has a long history. The Caesar cipher, which shifts every letter of the alphabet by a fixed number of places, goes back to Roman times and can be broken by hand in minutes. Modern symmetric algorithms such as AES and public-key systems such as RSA or elliptic-curve cryptography are in a different league altogether. The lesson of that history is that each generation of protection was eventually overtaken by the next generation of attack, and defenses had to evolve along with it.
What changes with AI
AI doesn't make 2256 any smaller. What it changes is everything around the mathematics. Real systems are rarely broken through their encryption. They are broken through a bug in the code, a misconfigured server, a reused password or a person who clicks the wrong link. Finding those weak spots used to take skilled people and a lot of time. AI makes the search faster and cheaper:
- Models can read large codebases and point to suspicious code paths far faster than a human reviewer. In late 2024, Google's Big Sleep project reported that an AI agent had found a previously unknown, exploitable bug in SQLite, one of the most widely deployed pieces of software in the world.
- Automated systems can run millions of attack simulations without tiring, trying variations no human team would ever have time for.
- Language models make phishing messages fluent and personal, which removes one of the oldest warning signs.
An analogy helps here. Prisons are built to hold an ordinary person: five-metre walls, iron bars, guards and cameras. Now imagine an inmate who can jump ten metres, or squeeze through gaps no human could fit through. The walls haven't changed, but they no longer do their job. Software is designed the same way, against the attacker its designers could imagine. In the digital world, AI is handing attackers something very much like the ability to jump ten metres.
AI on the defending side
The same capabilities work for defenders, and defense organisations are already using them:
- Finding your own weaknesses first. Automated code analysis and AI-driven testing can uncover vulnerabilities before an adversary does. DARPA's AI Cyber Challenge, which concluded in 2025, was built around exactly this idea: systems that find and patch bugs in open-source software on their own.
- Sensors and signals. Computer vision and deep learning models classify radar returns, satellite images and other sensor data far faster than human analysts.
- Network defense. Anomaly detection spots unusual traffic, and automated responses can isolate a compromised machine within seconds instead of hours.
A shifting balance of power
This is where the question stops being purely technical and becomes strategic. If AI can get around, outwit or quietly slip into complex defense systems, the balance of power changes. A country whose defenses can be turned against it can't really call itself strong. In that world, whoever controls the most capable AI may have more influence than whoever has the largest army.
The future may not bring people who can jump ten metres, but it will bring AI systems that multiply what both attackers and defenders can do. When it does, the idea of an "unbreakable software wall" will be little more than a comforting memory. Real security will depend on systems that detect and respond at machine speed, and on keeping people in charge of the decisions that matter.
