Cryptography protects data with mathematical problems that machine learning has not been shown to solve, so AI does not replace or defeat standard encryption. The two fields meet in narrower places: research on networks that learn to encrypt, deep-learning attacks on the physical leakage of a device running a cipher, and the planned move to quantum-resistant standards that NIST published on August 13, 2024. The quantum change, not AI, is what is reshaping encryption rules in Canada and the United States.
Key takeaways
- NIST released three post-quantum standards on August 13, 2024: FIPS 203 (ML-KEM), FIPS 204 (ML-DSA), and FIPS 205 (SLH-DSA).
- A 2016 Google Brain experiment trained neural networks to hide messages from a third network, but it was a research demonstration, not a usable cipher.
- Published deep-learning work on side-channel analysis attacks measurements from a device running AES, not the AES math itself.
- Canada's federal migration roadmap sets the end of 2031 for high-priority systems and the end of 2035 for the rest.
- This page is general information about security standards, not security, legal, or investment advice.
What encryption relies on, and what AI changes
Modern encryption depends on problems that are believed to be hard for any computer, such as recovering a secret key from public information. A machine learning model finds statistical patterns in examples. It performs well when a pattern exists and is poorly suited to a function designed to look random. That mismatch explains why the documented AI work in this area targets weak points around the cipher: how keys are exchanged, how software is built, and what a chip leaks while it computes.
This also frames how to read headlines about "AI-proof" or "AI-powered" encryption. Standards bodies approve algorithms after years of public analysis by cryptographers. A product that describes its encryption as designed by AI has not, for that reason alone, been through that process, and a reader can ask which published standard it implements.
A network that learned to hide messages
In an arXiv paper submitted on October 21, 2016, Martín Abadi and David Andersen of Google Brain described three neural networks named Alice, Bob, and Eve. Alice and Bob were trained to communicate so that Eve, an eavesdropper trained at the same time, could not recover the content. The authors did not prescribe an algorithm, and Alice and Bob learned a way to scramble and unscramble the data on their own.
The result is a study of learning in an adversarial setting. The test is confidentiality against one particular trained network, which is a narrower standard than a proof that no attacker could succeed. Nothing in the paper's abstract presents the learned scheme as a replacement for approved ciphers, and no standards body has adopted such a scheme.
Deep learning against a chip's leakage
A device that runs encryption gives off physical signals, such as electromagnetic emissions, whose pattern can depend on the secret key. Attacks that read those signals are called side-channel attacks. A 2018 paper in the IACR Cryptology ePrint Archive by Emmanuel Prouff, Rémi Strullu, Ryad Benadjila, Eleonora Cagli, and Cécile Dumas studied multilayer perceptrons and convolutional neural networks for this task. The authors released ASCAD, an open database built from measurements of a masked AES implementation, so that other researchers could repeat and compare results.
The target in that work is an implementation on hardware, not the AES algorithm as mathematics. Defenses exist at the implementation level, which is why the benchmark used a masked version designed to resist simpler attacks. For ordinary users of encrypted web or banking services, the practical concern in this literature is a lab setting with measurement access to a device.
Where each documented link sits
| Link between AI and cryptography | What the source documents | Limit on the claim |
|---|---|---|
| Networks learning to encrypt | Abadi and Andersen, arXiv, October 2016 | Tested against one trained eavesdropper, not a proof |
| Deep learning for side-channel analysis | Prouff and colleagues, IACR ePrint 2018/053 | Attacks device leakage on a masked AES implementation |
| Quantum-resistant standards | NIST FIPS 203, 204, and 205, August 13, 2024 | Driven by quantum computing, not AI |
| Migration schedules | NIST draft IR 8547 and the Government of Canada roadmap | Draft status for NIST; dates apply to government systems |
The quantum deadline that is driving change
NIST described FIPS 203 as a key-encapsulation standard, which lets two parties agree on a shared secret over a public channel, and FIPS 204 and FIPS 205 as digital signature standards. ML-KEM and ML-DSA rely on lattice problems, and SLH-DSA is built from hash functions. On March 11, 2025, NIST also selected a second key-encapsulation method, HQC, for standardization, with a draft standard still to come.
The reason is a possible future quantum computer. NIST's draft report IR 8547, published as an initial public draft on November 12, 2024, quotes National Security Memorandum 10, which sets 2035 as the goal for reducing quantum risk across federal systems. The draft table proposes that RSA and ECDSA at 112 bits of security strength be deprecated after 2030 and disallowed after 2035. The document is a draft, and a final version could not be confirmed for this page.
In Canada, the Government of Canada's post-quantum migration roadmap (ITSM.40.001), published June 23, 2025, asks federal departments to produce an initial migration plan in April 2026, finish high-priority systems by the end of 2031, and finish the remainder by the end of 2035.
Questions to ask about a product claiming AI-based security
- Which published standard or algorithm does it implement, and is that standard approved by NIST or by a Government of Canada cyber security body?
- Is the claim about the encryption itself, or about an AI feature that watches for threats around it?
- Has an independent party published an assessment, and does it cover the implementation as well as the design?
- What happens to stored data if the algorithm is later deprecated?
FAQ
Can AI break modern encryption?
None of the sources reviewed for this page shows AI breaking a standard cipher such as AES. Published deep-learning results concern physical leakage from a device, and researchers evaluate those on test setups.
Is encryption designed by an AI safe to use?
No standard exists for it. The 2016 experiment showed that networks can learn to hide data from another network, which is different from meeting the review that approved algorithms receive.
Do the new post-quantum standards mean quantum computers can already break RSA?
The documents cited here describe planning for a future machine of that capability. They set migration schedules so that systems are changed before it arrives.
Does this affect ordinary consumers?
Migration is mainly handled by software makers, browsers, and organizations that hold data. Individuals generally benefit by keeping devices and applications updated.
Related reading
- How predictive analytics uses data
- The components an AI system needs
- AI programs that sort through large data sets
- AI in research and discovery
More in AI in finance.
Last verified: October 2026
Sources
- NIST Releases First 3 Finalized Post-Quantum Encryption Standards, National Institute of Standards and Technology, August 13, 2024
- Learning to Protect Communications with Adversarial Neural Cryptography, Abadi M, Andersen D, arXiv:1610.06918, 2016
- Study of Deep Learning Techniques for Side-Channel Analysis and Introduction to ASCAD Database, Prouff E et al., IACR Cryptology ePrint Archive, 2018
- Transition to Post-Quantum Cryptography Standards (NIST IR 8547, initial public draft), NIST, November 12, 2024
- HQC announced as a 4th round selection, NIST, March 11, 2025
- Roadmap for the migration to post-quantum cryptography for the Government of Canada (ITSM.40.001), Government of Canada cyber security authority, June 2025