Why large language models won't break symmetric encryption
The rapid progress of large language models in writing code, solving mathematical problems and recognizing complex patterns has led some observers to worry these systems could eventually break modern symmetric encryption algorithms like AES. Cryptographers say that concern stems from a misunderstanding of encryption's mathematical foundation.
Symmetric encryption is a system in which the same key encrypts and decrypts data. The security of modern algorithms like AES-256 rests on the sheer number of possible key combinations — a number larger than the count of atoms in the observable universe.
Experts stress that an AI model cannot 'guess' or 'learn' this key, because the output of a well-designed encryption algorithm is statistically indistinguishable from random data. A pattern-recognition system cannot find a pattern where none exists.
This differs fundamentally from the tasks where AI models excel. Language models work by learning statistical regularities in training data; a cryptographically secure encryption output, by definition, contains no statistical regularities to learn.
Cryptographers note that even quantum computers offer only a limited advantage against symmetric encryption. A quantum method known as Grover's algorithm can reduce key-search time by a square-root factor, but that advantage is easily offset by simply doubling key length.
The piece argues the real threat lies not in mathematically breaking encryption but in implementation errors. Weak key generation, software bugs, side-channel attacks and human error have historically been the actual weak points of encryption systems.
AI models can genuinely help find these secondary weaknesses — for instance, by reviewing code to spot vulnerabilities or identifying common misconfigurations. But that is a fundamentally different capability from breaking the underlying algorithm.
Some researchers acknowledge AI could accelerate cryptanalysis research, helping human researchers discover new attack methods faster. That, however, is not the same as a 'magic shortcut' that would suddenly collapse the underlying mathematical security of existing algorithms.
Experts say part of the public confusion stems from AI's general capabilities being overstated; a system performing impressively on language and reasoning tasks does not mean it can exceed mathematically proven security boundaries.
Ultimately, the security community argues attention should focus on the real threat quantum computers pose to public-key encryption like RSA, and on the human-driven errors in how encryption systems are implemented — while AI breaking symmetric encryption itself appears scientifically implausible for the foreseeable future.
Read next
You can now turn off Google Gemini's visible AI watermarks. Here is what that actually changes
Google now lets users toggle off the visible watermark that normally appears on images, videos and music generated with Gemini and its Nano Banana and Omni models. The invisible SynthID watermark and embedded metadata remain in place regardless, meaning the content stays technically traceable as AI-generated even when it no longer looks that way.
A screen-sharing flaw is letting attackers take full control of Macs. Here is what to know
Security researchers have confirmed active exploitation of a vulnerability that lets remote attackers log into Macs without a password through a flaw in the screen-sharing feature. Here is how the exploit works, which systems are exposed, and the steps Mac users should take while a patch is developed.
OpenAI and Anthropic cut prices as Chinese AI rivals gain ground
OpenAI and Anthropic have both released cheaper models in recent weeks, a shift analysts attribute to intensifying competition from Chinese AI labs offering comparable performance at a fraction of the cost. The pricing pressure marks a notable change for two companies that have spent years competing primarily on capability rather than cost.
What we know about the alleged Iranian hacks on US water utilities
Over the past several weeks, hackers have broken into the industrial control systems of multiple US water utilities in an operation researchers attribute to actors linked to the Iranian government. Here is what has been confirmed about the intrusions, the vulnerable equipment involved, and why water infrastructure keeps turning up as a target.
Self-driving trucks are now testing on California highways: what the new permits actually allow
Aurora Innovation and Kodiak AI have received permits from the California Department of Motor Vehicles to test self-driving trucks on the state's highways, a milestone for an industry that has struggled to gain regulatory footing in one of the country's largest freight markets. Here is what the permits do and do not allow.