PQC & AI large models
Post-quantum cryptography and AI intersect in three distinct ways: AI helps find and fix cryptographic risk, PQC helps protect AI systems and their secrets, and the two combine in privacy-preserving machine learning. None of these means "AI broke PQC" — a caveat we return to at the end.
1. AI assisting cryptography & migration
Large language models are increasingly useful for the unglamorous work of crypto inventory and remediation. LLM-based crypto-misuse detection has been studied on real codebases (arXiv 2407.16576) where a model surfaced 63 new vulnerabilities — though roughly half were false positives, underscoring that human review remains mandatory. Tooling such as CryptoLLM and LLM-driven CBOM / crypto-asset discovery automates building the cryptographic bill of materials that any migration depends on.
Research is pushing toward end-to-end automation, combining PQC-aware detection, semantic refactoring, and hybrid verification, with LLM-assisted quantum-safe code auditing an active area of exploration.
On the commercial side, CBOM and crypto-discovery tools include SandboxAQ AQtive Guard, IBM Quantum Safe Explorer, and Keyfactor Command Risk Intelligence (Keyfactor acquired InfoSec Global in May 2025). See Discovery & inventory and Migration.
Important grounding: ML-aided cryptanalysis works only on reduced-round or toy ciphers (eprint 2025/1306). By contrast, ML-KEM and Dilithium have machine-checked security proofs (EasyCrypt / Jasmin, CRYPTO 2024) — the formal guarantees run the opposite direction of any AI "attack" hype.
2. Protecting AI with PQC
AI infrastructure is itself a HNDL target. The Cloud Security Alliance paper "Harvest Now, Decrypt Later: Quantum Risk to AI Infrastructure" recommends signing model artifacts and SBOMs with ML-DSA so provenance survives a quantum break. Per vendor materials/roadmaps, NVIDIA confidential computing (Hopper / Blackwell GPU TEE plus remote attestation) uses hybrid ML-KEM-1024 and is crypto-agile by design. PQC TLS protects AI and API traffic end to end, including Cloudflare's PQC SASE. Per vendor materials/roadmaps, Google's 2029 quantum-safe target explicitly includes Android ML-DSA. See Hybrid deployment.
3. PQC + privacy-preserving ML
Federated learning and other privacy-preserving ML schemes rely on cryptographic aggregation that must itself become quantum-safe. Post-quantum secure aggregation using a key-homomorphic PRF, and code-based aggregation, let a server combine client updates without learning individual contributions. Combinations of FHE (fully homomorphic encryption) with quantum federated learning are an active research frontier. See Resources.
Standards & references
PQC 与 AI 大模型
后量子密码与 AI 在三个层面交汇：AI 帮助发现并修复密码学风险，PQC 帮助保护 AI 系统及其密钥，二者又在隐私保护机器学习中结合。但这都不等于"AI 攻破了 PQC"，我们会在文末专门澄清这一点。
1 AI 辅助密码学与迁移
大模型在枯燥但关键的密码资产盘点与修复工作上越来越有用。基于 LLM 的密码误用检测已在真实代码库上研究过（arXiv 2407.16576）：模型发现了 63 个新漏洞，但约一半为误报，说明人工复核仍不可或缺。CryptoLLM 以及 LLM 驱动的 CBOM / 密码资产发现可自动构建迁移所依赖的密码物料清单。
相关研究正推动端到端自动化，结合 PQC 感知检测、语义重构与混合验证；LLM 辅助的抗量子代码审计也是活跃的探索方向。
商业侧的 CBOM 与密码发现工具包括 SandboxAQ AQtive Guard、IBM Quantum Safe Explorer 与 Keyfactor Command Risk Intelligence（Keyfactor 于 2025 年 5 月收购 InfoSec Global）。参见 资产发现与盘点 与 迁移。
重要澄清：ML 辅助的密码分析仅在缩减轮数或玩具级密码上有效（eprint 2025/1306）。相反，ML-KEM 与 Dilithium 拥有机器可验证的安全性证明（EasyCrypt / Jasmin，CRYPTO 2024），其形式化保证与任何 AI"攻破"的炒作方向恰恰相反。
2 用 PQC 保护 AI
AI 基础设施本身就是 HNDL 的目标。云安全联盟（CSA）报告《Harvest Now Decrypt Later 量子风险与 AI 基础设施》建议用 ML-DSA 对模型工件与 SBOM 签名，使来源可信在量子破解后依然成立。据厂商材料/公开路线图，NVIDIA 机密计算（Hopper / Blackwell GPU TEE 加远程证明）采用混合 ML-KEM-1024，并在设计上支持密码敏捷。PQC TLS 端到端保护 AI 与 API 流量，包括 Cloudflare PQC SASE。据厂商材料/公开路线图，Google 的 2029 抗量子目标明确包含 Android ML-DSA。参见 混合部署。
3 PQC 与隐私保护机器学习
联邦学习等隐私保护 ML 方案所依赖的密码聚合本身也必须抗量子。基于密钥同态 PRF 的后量子安全聚合，以及基于编码的聚合，让服务器在不获知单个客户端贡献的前提下合并更新。FHE（全同态加密）与量子联邦学习的结合是活跃的研究前沿。参见 资源链接。