P vs NP is one of those problems that sits quietly underneath a lot of modern computing.
The short version is that P is the class of problems we can solve efficiently. NP is the class of problems where we can verify a proposed solution efficiently. The big open question is whether every problem with an efficiently verifiable answer also has an efficiently discoverable answer.
That sounds abstract until you remember how much security depends on some things being easy to check and hard to find.
Cryptography is related to that gap, but not in a simple one-to-one way. Secure cryptography depends on stronger assumptions than P not equaling NP. It needs problems that are hard on average, not just hard in the worst case.
The AI Connection
The AI era does not mean P vs NP has been solved.
That is the wrong conclusion to jump to. Large language models and agentic systems are not magic complexity theory machines. They do not turn exponential problems into polynomial ones just because they are useful.
But they may still matter a lot.
AI can make search better. It can generate heuristics. It can write and test code quickly. It can compare strategies. It can help with theorem proving. It can read papers, reproduce experiments, and connect techniques across fields faster than a person working alone.
That does not prove a new complexity result. It does change the practical rate at which we explore hard problem spaces.
The distinction matters. A formal proof that P equals NP, paired with a practical algorithm, would be a civilizational event. A steady stream of better AI-assisted attacks against specific cryptographic systems would be quieter, but still operationally serious.
Cryptography Is An Assumption Stack
Mission-critical systems often rely on cryptographic algorithms as if they are permanent foundations.
They are not.
They are assumptions with engineering around them. We assume certain problems are hard enough. We assume implementations do not leak too much. We assume key management is disciplined. We assume attackers do not have a better shortcut than the public literature suggests.
AI puts pressure on that assumption stack.
Not because it automatically breaks RSA, elliptic curves, hashing, or symmetric encryption. The risk is more uneven than that. RSA and elliptic-curve systems depend on structured mathematical problems like factoring and discrete logarithms. Hashes and symmetric ciphers depend more on resistance to cryptanalysis and brute-force economics.
AI may help find implementation bugs. It may improve side-channel analysis. It may automate protocol review. It may discover better reductions, attacks, or parameter weaknesses. It may make expensive analysis cheap enough to run continuously.
That is still a real change.
Faster Problem Solving Changes Risk
The more problems we can solve faster and more accurately with AI, the more we should expect old security margins to be re-tested.
Some margins will hold. Some will not. The dangerous systems are the ones that cannot rotate algorithms, upgrade keys, or recover from a broken assumption.
Quantum computing is the clearer example of this pressure. Shor’s algorithm already gives us a theoretical break for RSA and elliptic-curve cryptography if large fault-tolerant quantum computers become practical. That is why post-quantum cryptography migration matters now, before the crisis is visible in production.
This is why cryptographic agility matters. Mission-critical systems should be designed so that algorithms can be replaced before a crisis. Keys should have lifetimes. Dependencies should be inventoried. Protocols should be documented well enough that a future team can understand what must change.
That is boring work until it suddenly is not.
The Practical Takeaway
P vs NP is still a mathematical question.
AI changes the engineering environment around it. It gives us faster tools for search, experimentation, verification, and attack development. That means cryptographic systems should be treated as living systems, not museum pieces.
The right posture is not panic. It is disciplined paranoia.
Know what algorithms you depend on. Know why you trust them. Know how you would replace them. Then assume the pace of discovery is going to keep increasing.