Visa Warns Intercepted AI Data Faces Retroactive Quantum Decryption

Protecting Enterprise AI Data in the Post-Quantum Era With Visa Inc — Ai4

Ai4

Post-quantum cryptography protects the trust infrastructure around AI pipelines rather than model weights, requiring immediate migration before quantum computers retroactively compromise harvested enterprise data.

AI-assisted summary of Protecting Enterprise AI Data in the Post-Quantum Era With Visa Inc — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the channel Ai4 on August 15, 2026, this presentation features Camille Nwaefulu, Senior Engineer in Applied and Post-Quantum Cryptography at Visa Inc. According to the session description, the talk examines post-quantum cryptography through an enterprise data-protection lens, guiding organizations accelerating AI adoption on how to protect high-value data over its confidentiality lifetime using risk prioritization, data classification, and crypto-agile designs.

## TL;DR Nwaefulu addresses the emerging threat of 'harvest now, decrypt later' attacks, where adversaries intercept and store encrypted enterprise AI data today to decrypt retroactively once cryptanalytically relevant quantum computers emerge. While symmetric ciphers like AES-256 retain quantum resilience, vulnerable public key algorithms like RSA and ECC must systematically migrate to finalized NIST standards such as ML-KEM, ML-DSA, and SLH-DSA. Emphasizing that post-quantum cryptography protects the surrounding trust infrastructure rather than the AI models themselves, she outlines a structured transition framework focused on data classification, confidentiality lifespans, crypto-agility, and pilot testing.

## Atmosphere & style The presentation maintains a pragmatic, measured, and engineering-focused tone, opening with an informal, self-deprecating comment about jet lag before moving briskly into a clear-eyed security briefing. Nwaefulu deliberately strips away the speculative sci-fi mystique of quantum computing, grounding the topic in practical enterprise architecture, clear schematics, and realistic risk management. The session feels structured, steady, and accessible, reassuring practitioners that mitigating quantum risks requires methodical preparation and cross-functional discipline rather than quantum hardware.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=4s) Adversaries are actively executing 'harvest now, decrypt later' (HNDL) attacks by capturing encrypted enterprise and AI data today to decrypt it retroactively once cryptanalytically relevant quantum computers exist. - [02:59](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=179s) Post-quantum migration is fundamentally an enterprise-wide risk management discipline structured across four operational phases: See the Risk, Set Priority, Build Readiness, and Make Transition. - [05:09](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=309s) Quantum computing threatens public key algorithms like RSA and ECC via Shor's algorithm, while symmetric ciphers such as AES-256 remain quantum-resilient against Grover's quadratic speedup. - [05:09](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=309s) Organizations must measure their quantum risk by evaluating their required data confidentiality lifetime against their projected cryptographic migration timeline. - [07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) Post-quantum cryptography executes on standard classical commodity hardware and servers, using complex mathematical structures like lattice problems rather than requiring quantum computers. - [07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) NIST has finalized core post-quantum standards: ML-KEM for key encapsulation, ML-DSA for digital signatures, and SLH-DSA for stateless hash-based authentication. - [09:05](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=545s) The end-to-end enterprise AI pipeline exposes multiple cryptographic touchpoints, spanning identity gateways, RAG vector databases, model inference engines, logging telemetry, and external APIs. - [12:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=727s) Post-quantum cryptography does not alter AI model weights directly; it secures the surrounding trust infrastructure, including TLS connections, PKI certificates, workload identities, and model signing pipelines. - [14:34](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=874s) Migration efforts should be prioritized objectively by classifying data sensitivity, confidentiality lifetime, cryptographic exposure, and legacy dependency complexity before rolling out to production. - [17:05](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1025s) Enterprises must avoid common pitfalls such as treating PQC as a simple drop-in swap, ignoring operational overheads like larger key sizes and latency, and neglecting foundational entropy. - [18:34](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1114s) Post-quantum transition roles vary by scale: small businesses must audit vendor security and key management roadmaps, while large enterprises require broad cross-functional alignment and governance. - [20:15](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1215s) A structured 90-day preparation roadmap divided into 30-, 60-, and 90-day milestones enables organizations to discover assets, assign ownership, and pilot crypto-agile architectures effectively.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=4s) Introduction to Enterprise AI and "Harvest Now, Decrypt Later"

Camille Nwaefulu, a Senior Cybersecurity Engineer specializing in applied and post-quantum cryptography at Visa Inc., opens the presentation with an informal greeting, joking about being on Eastern Time and feeling "all over the place." The central focus of the talk is examining how enterprises must protect artificial intelligence data at scale by adopting post-quantum cryptography (PQC). As enterprises rapidly deploy AI systems, sensitive corporate data is increasingly collected, aggregated, and routed across diverse cloud environments, APIs, and vector databases, creating an expanded attack surface where classical cryptographic assumptions no longer hold.

The most urgent vector discussed is the strategy known as "harvest now, decrypt later" (HNDL). Adversaries are actively intercepting and storing encrypted enterprise network traffic today, even if they cannot break the encryption immediately. These bad actors intend to warehouse this encrypted ciphertext for five, ten, or fifteen years until fault-tolerant quantum computers emerge. Once cryptanalytically relevant quantum computers are operational, that preserved data will be retroactively decrypted. For enterprises holding intellectual property, proprietary algorithms, pending patents, or long-term confidential records, data captured today will become completely compromised in the future if relying solely on quantum-vulnerable public key algorithms.

```mermaid flowchart LR accTitle: Harvest Now Decrypt Later Attack Timeline subgraph Present["Present Day"] A["Data Transmission<br/>(RSA / ECC Encrypted)"] --> B["Adversary Interception<br/>& Indefinite Storage"] end subgraph Future["Quantum Era (5-15 Years)"] B --> C["Cryptanalytically Relevant<br/>Quantum Computer (CRQC)"] C --> D["Retroactive Plaintext<br/>Decryption & Exploitation"] end ```

- Adversaries target long-term confidential assets such as secret patents, intellectual property, and proprietary research. - Inter-data-center traffic and cloud migrations present prime harvesting opportunities for persistent surveillance. - Classical symmetric encryption (e.g., AES-128/256) and public key mechanisms (e.g., RSA, ECC) secure communications today, but the mathematical underpinnings of public key trust establishment remain exposed to future quantum attacks.

---

[02:59](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=179s) A Four-Phase Roadmap: Turning Quantum Risk into Action

To help organizations navigate the post-quantum shift without becoming overwhelmed, Nwaefulu breaks down the enterprise transition into four structured operational phases. The initial challenge is recognizing that post-quantum migration is fundamentally an enterprise-wide risk management discipline rather than an isolated IT patch.

```mermaid flowchart LR accTitle: Four Phases of Quantum Risk Remediation P1["Phase 1: See the Risk<br/>• Quantum Threat Analysis<br/>• Vulnerability Mapping"] --> P2["Phase 2: Set Priority<br/>• Data Classification<br/>• Resource Allocation"] P2 --> P3["Phase 3: Build Readiness<br/>• Crypto-Agile Design<br/>• Hardware Compatibility"] P3 --> P4["Phase 4: Make Transition<br/>• Standards Migration<br/>• Cross-Functional Execution"] ```

1. **Phase 1: See the Risk** — Execute a thorough quantum threat analysis across the entire organization to pinpoint precisely where quantum-vulnerable cryptographic algorithms and dependencies reside. 2. **Phase 2: Set the Priority** — Classify enterprise data assets and identify high-risk repositories. Enterprises have finite budgets and engineering bandwidth, so funding and manpower must be allocated toward protecting high-value assets—such as healthcare records and trade secrets—rather than public or low-consequence data. 3. **Phase 3: Build Readiness** — Architect crypto-agile infrastructures that can swap cryptographic primitives without breaking downstream business applications or hardware security modules (HSMs). 4. **Phase 4: Make the Transition** — Systematically deprecate quantum-vulnerable public key algorithms in favor of finalized PQC standards, assigning explicit operational responsibilities across cross-functional groups, including technical, executive, and legal teams.

---

[05:09](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=309s) The Quantum Threat: Public Key Vulnerability vs. Symmetric Resilience

The impact of quantum computing is asymmetric across different cryptographic primitives. The primary and immediate target of a cryptanalytically relevant quantum computer is public key cryptography. Algorithms like RSA and Elliptic Curve Cryptography (ECC) rely on specific mathematical problems—namely integer factorization and discrete logarithms—which Shor's algorithm running on a sufficiently large quantum computer can solve in polynomial time. This breaks foundational security layers including digital signatures, identity certificates, mutual TLS handshakes, and key establishment mechanisms.

In contrast, symmetric key cryptography (such as AES-256) is not vulnerable in the same fundamental way. Grover's algorithm provides a quadratic quantum speedup against symmetric ciphers, effectively halving key strength (e.g., reducing AES-256 to 128 bits of quantum security), meaning strong symmetric encryption remains a robust component of post-quantum security architectures. When calculating quantum risk, organizations must evaluate their *confidentiality lifetime* against their projected *migration timeline*. If an organization requires confidential data to remain protected for ten years, but migrating legacy systems takes five years and quantum computing maturity is expected within that combined window, data protection is already failing today.

| Cryptographic Type | Common Algorithms | Primary Use Cases | Quantum Vulnerability & Impact | Post-Quantum Posture | |:--- |:--- |:--- |:--- |:--- | | **Public Key Cryptography** | RSA, ECC (ECDH, ECDSA) | Key establishment, digital signatures, certificates, zero-trust identity | **Critically Vulnerable**: Polynomial-time break via Shor's algorithm; destroys identity trust and key exchange. | Must be systematically migrated to NIST PQC standards (e.g., ML-KEM, ML-DSA). | | **Symmetric Cryptography** | AES-128, AES-256 | Bulk data-at-rest and data-in-transit payload encryption | **Resilient**: Quadratic speedup via Grover's algorithm; effectively halves key length. | AES-256 remains secure and integral to PQC hybrid architectures. |

---

[07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) Demystifying Post-Quantum Cryptography and NIST Standards

A frequent industry misunderstanding is that post-quantum cryptography requires quantum hardware to operate. Nwaefulu explicitly clarifies: "Post-quantum cryptography does not require quantum computers." PQC algorithms are classical software routines engineered to execute on existing commodity laptops, enterprise servers, microservices, and standard cloud infrastructures.

The innovation of PQC lies entirely in the underlying mathematics. Rather than depending on number-theoretic problems that quantum processors easily solve, PQC relies on complex mathematical structures, primarily lattice-based problems, which are believed to be computationally intractable for both classical and quantum computers.

Nwaefulu highlights the three primary standards finalized by the National Institute of Standards and Technology (NIST):

- **ML-KEM (Module-Lattice-Based Key-Encapsulation Mechanism)**: Formerly known as **Kyber**, this standard functions as the secure handshake for establishing shared encryption keys across untrusted networks. - **ML-DSA (Module-Lattice-Based Digital Signature Algorithm)**: Formerly known as **Dilithium**, this provides the "digital seal" used to sign code, authenticate servers, and validate digital identity certificates. - **SLH-DSA (Stateless Hash-Based Digital Signature Algorithm)**: Formerly known as **SPHINCS+**, this algorithm serves as a non-lattice, hash-based alternative for digital signatures, providing defense-in-depth in the event of unforeseen mathematical breakthroughs against lattice structures.

---

[09:05](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=545s) The Enterprise AI Stack and Exposure Points

Protecting an AI deployment requires mapping data flows across every step of the end-to-end AI operational lifecycle. When an employee or customer submits a prompt, it triggers a chain of interactions across distinct components that each rely on cryptographic trust.

```mermaid flowchart TD accTitle: Enterprise AI Data Flow and Security Controls User["User / Client Application"] -->|"1. Authenticated Prompt"| IDM["Identity Management & Gateways"] IDM -->|"2. Rate-Limited API Request"| Ingest["Data Ingestion & APIs"] Ingest -->|"3. Semantic Retrieval Query"| VDB["Enterprise Vector Databases & Data Stores"] VDB -->|"4. Grounded Context"| Model["Foundation Model / Inference Engine"] Model -->|"5. Generated Inference"| Log["Monitoring, Logging & Audit Systems"] Log -->|"6. Failover & Storage"| Backup["Backups, Long-Term Retention & 3rd-Party APIs"] ```

1. **Identity Management**: Validates user credentials, role-based access controls, and session permissions before transactions begin. 2. **AI Gateway**: Serves as the central gatekeeper enforcing rate limits, policy constraints, and security inspection. 3. **Enterprise Data Sources & Vector Databases**: Stores internal intellectual property and powers semantic retrieval (Retrieval-Augmented Generation / RAG) to supply contextual grounding. 4. **Foundation Models & Inference Engines**: Executes inference over incoming prompts, model weights, and retrieved context. 5. **Monitoring, Logging, and Observability**: Records inputs, outputs, operational telemetry, and feedback loops to guide future model optimization. 6. **Retention, Backups, and Third-Party Integrations**: Manages cold storage, rollback snapshots, and outbound API calls to external vendors.

Data vulnerabilities emerge throughout this chain: during raw data ingestion, within semantic embedding lookups, across service-to-service microservice APIs, around stored model weights, and within external third-party integrations. If third-party suppliers lack a defined PQC migration roadmap, they introduce blind spots into the host organization's security posture.

---

[12:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=727s) Securing the Trust Infrastructure Around AI

Nwaefulu addresses a major misconception in enterprise security: "People believe that post-quantum cryptography is going to protect the AI model itself. And that is not exactly the correct case. Instead, it's going to protect the trust infrastructure around the actual artificial intelligence."

PQC does not alter the deep learning matrix multiplications or model weights directly; rather, it fortifies the surrounding fabric that guarantees authenticity, confidentiality, and integrity. This includes:

- **Quantum-Resistant Connections**: Deploying post-quantum TLS handshakes across AI gateways, vector databases, and internal API communication channels. - **Quantum-Resistant Identity and Trust**: Upgrading PKI certificates, mutual TLS (mTLS), and workload identities to prevent adversaries from forging credentials. - **Trusted Software Models**: Cryptographically signing model artifacts, fine-tuning checkpoints, and deployment pipelines using post-quantum digital signatures (such as ML-DSA). - **Long-Term Data Protection**: Securing stored prompt logs, knowledge bases, and vector embeddings against HNDL exploitation.

Risk profiles differ substantially across organizational data tiers. High-priority assets like proprietary source code, customer financial records, and medical data demand immediate post-quantum safeguards due to their long shelf-life. Conversely, ephemeral data and public marketing copy do not justify the same capital expenditure or engineering allocation.

---

[14:34](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=874s) Practical Prioritization Framework and Crypto Agility

Organizations need an objective framework to assess and prioritize their cryptographic migration pipelines rather than attempting a high-risk, all-at-once transition:

```mermaid flowchart TD accTitle: Practical Prioritization and Assessment Framework subgraph Assessment["1. Evaluation Criteria"] C1["Data Sensitivity<br/>(Restricted, Confidential, Public)"] C2["Confidentiality Lifetime<br/>(Days vs. 5-15+ Years)"] C3["Cryptographic Exposure<br/>(Current Algorithms & Key Lengths)"] C4["Migration Complexity<br/>(Legacy Systems & Dependencies)"] end subgraph Execution["2. Staged Implementation"] E1["Migrate High-Value Sandboxes / Dev Labs"] E2["Verify Interoperability & System Readiness"] E3["Document Residual Risks & Roll Out to Production"] end Assessment --> Execution ```

- **Data Sensitivity and Classification**: Catalog all enterprise data into strict classifications (e.g., Restricted PII, Highly Confidential, Internal, Public) to avoid misallocating funds. - **Confidentiality Lifetime**: Calculate how long data must legally or competitively remain secret (e.g., 2 days versus 15 years). - **Cryptographic Exposure**: Audit existing production environments to determine what cipher suites, certificate authorities, and key lengths are actively deployed. - **Migration Complexity**: Quantify technical dependencies, engineering hours, and operational friction before beginning upgrades. - **Staged Lab Deployments**: Test and validate PQC algorithms in development environments, sandbox clusters, and pilot labs prior to production rollouts to verify interoperability and prevent service outages.

---

[17:05](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1025s) Entropy, Trust Foundations, and Common Migration Pitfalls

The security of post-quantum cryptography relies on the integrity of underlying operational primitives. Nwaefulu highlights entropy as a crucial foundation: without high-quality randomness, robust key management lifecycles, and hardened operational environments, even mathematically sound PQC algorithms can fail.

Enterprises should structure their long-term cryptographic strategy around a six-step lifecycle:

$$\text{Discover} \longrightarrow \text{Classify} \longrightarrow \text{Prioritize} \longrightarrow \text{Design} \longrightarrow \text{Pilot} \longrightarrow \text{Scale \& Govern}$$

Nwaefulu concludes this segment by warning against common migration pitfalls:

- **Treating PQC as a Simple Algorithm Swap**: Assuming that post-quantum ciphers can be dropped into existing codebases without architectural adjustments. - **Operating Without Visibility**: Attempting remediation without an exhaustive, automated cryptographic inventory of all internal systems and third-party dependencies. - **Perimeter-Only Protection**: Securing external load balancers while leaving internal microservice-to-microservice APIs vulnerable to internal harvesting. - **Ignoring Performance Overheads**: Overlooking the operational impact of PQC's significantly larger key sizes, digital signatures, and handshake latency on legacy network stacks and memory-constrained devices.

[18:34](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1114s) Cryptographic Migration Pitfalls and Organizational Roles

When planning post-quantum migrations, organizations must avoid common architectural traps. Changes in cryptographic key and ciphertext sizes must be evaluated so they "do not interrupt your models and your infrastructure." Other frequent pitfalls include overlooking underlying trust foundations, testing new algorithms in complete isolation from production realities, and attempting to migrate systems without designing for crypto-agility from the start.

```mermaid flowchart TD accTitle: Post-Quantum Transition Roles subgraph SmallBiz["Small Business Owners"] direction TB SB1["Audit 3rd-Party PQC Roadmaps"] SB2["Verify Key Storage & HSMs"] SB3["Check Client vs. Server Encryption"] end subgraph Enterprise["Enterprise Leaders"] direction TB EL1["Establish Governance Policies"] EL2["Align Cross-Functional Teams"] EL3["Coordinate Testing & Documentation"] end subgraph Technical["Technical Teams"] direction TB TT1["Take Cryptographic Inventory"] TT2["Audit All Critical Data Stores"] end ```

Every stakeholder has a distinct responsibility in the post-quantum transition depending on organizational scale: - **Small Business Owners**: Because smaller businesses rely heavily on external vendors and lack direct control over foundational infrastructure, owners must thoroughly evaluate their third-party vendors' PQC roadmaps. They must ask explicit questions about how data is secured, how cryptographic keys are stored, whether hardware security modules (HSMs) are deployed, and whether data relies on client-side or server-side encryption. - **Enterprise Leaders**: Large enterprises hold the resources to drive comprehensive post-quantum transformations. Leaders must establish cross-cutting governance policies and bring technical teams, product managers, non-technical stakeholders, technical writers, and QA testers into alignment. While deploying post-quantum algorithms can be challenging, broad cross-functional alignment ensures organizations "get further together." - **Technical Teams**: Engineers and practitioners must lead the discovery phase by cataloging cryptographic assets, taking full inventories of active algorithms, and auditing all critical data stores.

---

[20:15](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1215s) The 90-Day Preparation Roadmap and Strategic Next Steps

Moving an entire enterprise infrastructure over to post-quantum algorithms inevitably takes much longer than three months, but establishing a structured 90-day onboarding plan gives teams a clear runway. Dividing early efforts into distinct 30-day, 60-day, and 90-day milestones allows organizations to assess risk, organize necessary resources, and draft migration plans well before attempting full system integration.

- **Phase Breakdown**: Use 30-, 60-, and 90-day intervals to prepare environments, assign team ownership, and assemble technical documentation. - **Strategic Pillars**: Organizations can use standardized strategic pillars to dive deeper into crypto-agility, legacy asset mitigation, and vendor assurance frameworks. - **Closing**: The session wraps with a summary slide of core strategic pillars for attendees to capture and reference as they implement their post-quantum AI data security initiatives.

## Notable quotes > [07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) Post-quantum cryptography does not require quantum computers. — Camille Nwaefulu

> [12:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=727s) People believe that post-quantum cryptography is going to protect the AI model itself. And that is not exactly the correct case. Instead, it's going to protect the trust infrastructure around the actual artificial intelligence. — Camille Nwaefulu

> [18:34](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1114s) do not interrupt your models and your infrastructure. — Camille Nwaefulu

> [18:34](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1114s) get further together. — Camille Nwaefulu

## Mentioned Camille Nwaefulu, Visa Inc., Ai4, NIST, ML-KEM, Kyber, ML-DSA, Dilithium, SLH-DSA, SPHINCS+, RSA, ECC, ECDH, ECDSA, AES-128, AES-256, Shor's algorithm, Grover's algorithm

## Video description Camille Nwaefulu - Senior Engineer, Applied and Post-Quantum Cryptography, Visa Inc.

As enterprises accelerate AI adoption, they are collecting, combining, and moving greater volumes of sensitive data across increasingly complex environments. This talk will examine post-quantum cryptography through a data-protection lens, exploring why organizations must consider not only how data is secured today, but how long that protection must remain effective. I will discuss the practical challenges enterprises face, including identifying high-value data, understanding where vulnerable cryptography is used, managing legacy systems, and evaluating vendor readiness. Attendees will learn how data classification, risk-based prioritization, small pilot initiatives, and crypto-agile design can support a realistic post-quantum roadmap. This session will help practitioners connect AI growth, long-term data exposure, and post-quantum preparation without requiring a deep background in quantum computing.

[Source](https://www.youtube.com/watch?v=-cZ1lkUJL6Y)

中文简报

Visa警告：被截获的AI数据面临追溯性量子解密风险 后量子密码学保护的是AI流水线周围的信任基础设施而非模型权重，要求企业在量子计算机追溯性破解已采集数据前立即实施迁移。 Nwaefulu指出了“先收集，后解密”这一新兴威胁：攻击者当前截获并存储加密的企业AI数据，待具备密码分析能力的量子计算机出现后再进行追溯性解密。尽管AES-256等对称密码仍具备抗量子能力，但RSA和ECC等脆弱的公钥算法必须系统性迁移至已敲定的NIST标准（如ML-KEM、ML-DSA和SLH-DSA）。她强调后量子密码学保护的是周边的信任基础设施而非AI模型本身，并概述了一个聚焦于数据分类、保密期限、加密敏捷性和试点测试的结构化转型框架。 攻击者正在积极实施“先收集，后解密”（HNDL）攻击，通过在当下截获加密的企业及AI数据，以便在具备密码分析能力的量子计算机问世后实施追溯性解密。 (0:04) 后量子迁移在本质上是一项全企业范围的风险管理实践，具体分为四个运营阶段：洞察风险、设定优先级、构建就绪能力以及实施转型。 (2:59) 量子计算通过Shor算法对RSA和ECC等公钥算法构成威胁，而AES-256等对称密码在面对Grover算法的二次加速时仍保持抗量子能力。 (5:09) 组织必须通过比对其所需的数据保密期限与预估的密码算法迁移时间线，来评估自身的量子风险。 (5:09) 后量子密码学运行在标准的经典商用硬件和服务器上，采用格问题等复杂数学结构，并不需要量子计算机。 (7:17)

Reading summary

Camille Nwaefulu, a Senior Cybersecurity Engineer at Visa Inc., argues that enterprises scaling artificial intelligence must overhaul their supporting trust infrastructures using post-quantum cryptography. With adversaries actively intercepting encrypted enterprise traffic for future decryption, organizations must treat post-quantum migration as a proactive, risk-based discipline rather than an isolated software patch. Quantum Decryption and The Harvesting Threat (0:04) Nwaefulu warns that adversaries are actively executing "harvest now, decrypt later" (HNDL) attacks by capturing and storing encrypted enterprise network traffic today. These threat actors intend to warehouse ciphertext until cryptanalytically relevant quantum computers emerge—projected within five to fifteen years—at which point the data can be retroactively decrypted. This creates immediate exposure for long-lived sensitive assets such as intellectual property, secret patents, and confidential enterprise records transferred across cloud environments and vector databases. Asymmetric Cryptographic Risk and Mathematical Defenses (5:09) The threat posed by quantum computing is asymmetric across existing cryptographic primitives. Public-key mechanisms such as RSA and Elliptic Curve Cryptography rely on integer factorization and discrete logarithms, which Shor's algorithm can solve in polynomial time to destroy identity certificates, digital signatures, and TLS handshakes. In contrast, symmetric ciphers like AES-256 remain robust; Grover's algorithm provides only a quadratic quantum speedup, halving key strength while leaving 128 bits of post-quantum security. Nwaefulu clarifies that post-quantum cryptography (PQC) does not require quantum hardware, running instead on classical commodity servers and microservices. Finalized National Institute of Standards and Technology (NIST) algorithms address these vulnerabilities using complex mathematical structures like lattices. Key standards include ML-KEM (formerly Kyber) for shared key exchange, ML-DSA (formerly Dilithium) for digital signatures, and SLH-DSA (formerly SPHINCS+) as a stateless, hash-based alternative that offers defense-in-depth against unforeseen mathematical breaks in lattice theory. Securing Enterprise AI Trust Infrastructure (12:07) Addressing a common enterprise misconception, Nwaefulu highlights that post-quantum cryptography does not alter AI model weights or deep-learning matrix calculations directly. Instead, PQC secures the operational trust infrastructure surrounding artificial intelligence workflows, including user identity management, rate-limiting gateways, semantic retrieval queries in vector databases, model deployment pipelines, and telemetry audit logs. Securing the end-to-end AI lifecycle requires deploying post-quantum TLS across internal service APIs, implementing quantum-resistant mutual TLS (mTLS) to prevent credential forgery, and signing software models and fine-tuning checkpoints with algorithms like ML-DSA. Nwaefulu notes that because vector databases store corporate intellectual property to ground inference models, failing to protect inter-service communications leaves underlying knowledge bases exposed to data harvesting. Migration Roadmaps, Prioritization, and Pitfalls (14:34) Rather than attempting an abrupt migration, enterprises must prioritize workloads by cross-referencing data sensitivity, migration complexity, cryptographic exposure, and confidentiality lifetime. If an enterprise requires data to remain confidential for ten years and system migration requires five years, relying on quantum-vulnerable cryptography poses an active breach risk if quantum computers mature within that combined window. Assets with long retention requirements, such as trade secrets and medical data, require immediate protection over low-consequence or ephemeral data. Nwaefulu cautions against major migration pitfalls, including treating PQC as a simple drop-in algorithm swap, securing only perimeter networks while ignoring internal microservices, and neglecting the latency and memory overheads introduced by larger post-quantum keys and signatures. Organizations should follow a structured lifecycle—Discover, Classify, Prioritize, Design, Pilot, and Scale & Govern—starting with automated cryptographic inventories and 90-day preparation roadmaps. While enterprises coordinate cross-functional teams, smaller organizations must audit external vendor roadmaps regarding key storage, hardware security modules, and encryption tiers. Nwaefulu emphasizes that because classical public-key infrastructure will inevitably be compromised, enterprise data protection depends on building crypto-agile architectures today. The long-term challenge centers on managing implementation overheads and third-party dependencies before the window between system migration time and quantum availability closes.

精炼阅读摘要

Visa公司应用与后量子密码学资深网络安全工程师Camille Nwaefulu指出，针对企业AI数据的“先拦截后解密”（HNDL）威胁已在当下发生，企业亟须部署后量子密码学（PQC）以重塑AI外围的信任基础设施，而非仅关注模型算法本身。 “先拦截后解密”与公钥密码的量子危机 (0:04) Nwaefulu在演讲中指出，攻击者当前正主动拦截并持久化存储企业的加密网络流量，其策略是在未来5至15年具备容错能力的实用量子计算机（CRQC）出现时，再对这些密文进行追溯性解密，即“先拦截后解密”（HNDL）。对于持有专有算法、未决专利和商业秘密等长期机密资产的企业而言，若继续单纯依赖传统公钥加密算法，当前传输并被截获的数据在未来将面临全面被破解的确定性风险。 量子计算对经典密码体系的破坏呈现出显著的不对称性。Nwaefulu解释称，基于大整数分解与离散对数数学难解性的经典公钥密码体系（如RSA和ECC）极度脆弱，运行Shor算法的大规模量子计算机可在多项式时间内将其攻破，进而瓦解数字签名、公钥证书和TLS密钥协商机制。相反，以AES-256为代表的对称加密具有较强抗量子韧性，Grover算法仅能提供二次方加速，相当于将密钥安全强度减半至128位，因此AES-256在后量子时代依然安全有效。企业在评估量子风险时，必须将数据的保密生命周期与系统迁移耗时相加；若该总年限落入量子计算成熟的时间窗口内，现有数据保护实际上已经失效。 基于经典硬件运行的NIST后量子标准 (7:17) 针对产业界常见的认知误区，Nwaefulu明确澄清：后量子密码学并不依赖量子计算机硬件来运行。PQC本质上是经过工程化设计的经典软件例程，能够直接部署在现有的通用笔记本电脑、企业服务器、微服务架构以及标准云基础设施之上。PQC的防御有效性完全建立在底层数学结构的突破上——它放弃了易被量子处理器攻破的数论难题，主要转向以格理论（lattice-based）为核心的高维数学构造，此类难题被普遍认为对经典计算与量子计算均具备极高的计算不可行性。 演讲详细介绍了美国国家标准与技术研究院（NIST）敲定的三项核心PQC标准：首先是基于模块格的密钥封装机制ML-KEM（原名Kyber），负责在不可信网络中建立安全的共享对称密钥；其次是基于模块格的数字签名算法ML-DSA（原名Dilithium），充当代码签名、服务器身份鉴别和数字证书验证的数字印章；最后是无状态基于哈希的数字签名算法SLH-DSA（原名SPHINCS+），作为非格结构的哈希签名备选方案，可在未来格密码发生未预见性数学突破时提供纵深防御保障。 防护重心聚焦于AI外围信任基础设施 (12:07) 针对企业安全认知的另一大误区，Nwaefulu强调，PQC的目的并非直接保护人工智能模型本身的权重矩阵乘法运算，而是加固环绕在AI周围的信任基础设施。现代企业AI系统的交互链条十分复杂：用户输入提示词后，请求依次流经身份管理系统、AI安全网关、支撑检索增强生成（RAG）的企业向量数据库、基础大模型推理引擎，最终沉淀至监控日志与第三方集成存储中。在数据摄取、语义嵌入比对、微服务API通信等环节，处处依赖密码学建立信任链条。 构建抗量子AI安全防护需要多层次落地：在AI网关、向量数据库及内部微服务API之间部署后量子TLS连接；将公钥基础设施（PKI）证书与相互传输层安全（mTLS）工作负载标识升级为抗量子认证体系；运用ML-DSA等后量子数字签名对模型构件、微调检查点及CI/CD部署管道进行签名防篡改；并对长期留存的向量数据库知识库与提示词日志提供防范HNDL的长期保密保护。同时，企业应根据数据资产分类实施差异化防护，优先保障专有源代码、财务记录和医疗数据等高价值资产，避免在公开营销文案等瞬时数据上耗费同等成本。 敏捷治理框架与落地实施陷阱防范 (2:59) Nwaefulu将后量子迁移定义为企业级的整体风险管理工程，而非孤立的IT补丁修补，并提出了四阶段实施路径：“看清风险”（全量威胁分析与漏洞测绘）、“确立优先级”（资产分类与资源分配）、“构建就绪”（密码敏捷性设计与硬件兼容适配）以及“实施过渡”（标准迁移与跨职能落地）。企业应遵循“发现、分类、定级、设计、试点、规模化与治理”的六步生命周期，依托数据敏感度、保密期限、现有密码暴露面与系统迁移复杂度等维度评估优先级，并在沙箱实验室中率先验证互操作性以避免服务中断。 在执行细节上，Nwaefulu警告了一系列典型实施陷阱：切勿将PQC视为可直接替换的插件，后量子算法生成的密钥和签名体积显著增大，必须评估其对现有网络延迟、吞吐及受限硬件内存的性能冲击；严守高质量随机数（熵源）与密钥生命周期管理底线，防止因基础信任缺陷削弱算法安全性；杜绝仅防护外围负载均衡器而放任内部微服务裸露的边缘防御主义。不同角色需各司其职：小型企业应全面审查第三方供应商的PQC路线图与HSM存储规范；大型企业管理者需推动技术、法务、测试跨职能对齐；技术团队则应制定“30天、60天、90天”的阶段性路线图，分步开展全量密码资产盘点与落地演练。 Nwaefulu总结认为，尽管企业向后量子密码体系的全面过渡漫长且复杂，但面对当下正在发生的长期流量窃取风险，企业不能推迟行动窗口。唯有及早摸清内部资产底数、建立以密码敏捷性为核心的跨职能治理体系，并按优先级逐步加固高价值AI数据链路，才能在实用量子计算机真正到来时化被动为主动。

Detailed notes — zh

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 15 日发布，主讲人为来自 Visa 的应用与后量子密码学资深网络安全工程师 Camille Nwaefulu。演讲围绕数据保护视角展开，探讨在企业加速采用 AI 的背景下，如何应对未来量子计算对敏感数据的长期威胁，并介绍数据分级、风险优先级评估、试点落地以及密码敏捷性等实用应对框架。 ## TL;DR 演讲重点剖析了“先窃取，后解密”（HNDL）这一新型威胁模型，指出黑客正在拦截并长期保存加密流量，等待未来容错量子计算机问世后破解。面对公钥密码体系（RSA、ECC）即将失效的危机，主讲人澄清 PQC 本身只需运行在现有经典算力上，其核心使命是保护 AI 架构周边的传输链路、身份认证与信任基础设施，而非模型参数本身。企业需建立包含“识别风险、定级优先级、就绪准备、平稳迁移”的四阶段实施路径，警惕将 PQC 误判为简单“算法替换”的陷阱，并利用 90 天前瞻规划扎实推进加密资产摸底与供应链协同。 ## Atmosphere & style 演讲整体呈现出沉稳严谨且极具实践导向的专业技术氛围。主讲人以轻松幽默的倒时差开场打破严肃壁垒，随后迅速切换为条理清晰、抽丝剥茧的工程教学风格。讲解避开了高深晦涩的纯量子理论推导，始终聚焦于企业真实系统架构落地中的工程痛点与认知陷阱，给人一种兼具前瞻战略高度与踏实运维掌控力的理性说服感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=4s) 企业 AI 面临“先窃取，后解密”（HNDL）威胁，攻击者正在截获并存储加密通信流量，等待未来容错量子计算机成熟后解密历史敏感资产。 - [03:02](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=182s) 企业应对量子威胁应遵循四大阶段框架：识别风险、确定防护优先级、建立敏捷密码（Crypto-Agile）就绪能力以及跨部门协同推进迁移。 - [05:10](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=310s) 量子计算主要击穿依赖传统数论难题的公钥密码体系（如 RSA、ECC），而强对称加密算法（如 AES）在后量子架构中依然保持高度安全并发挥关键作用。 - [06:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=367s) HNDL 防护规划需权衡数据保密寿命与迁移周期，重点保护生命周期达数十年且具备长期战略价值的核心资产，避免在短期数据上过度消耗资源。 - [07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) 后量子密码学（PQC）无需量子计算机即可在现有服务器和通用基础设施上运行，NIST 已正式标准化 ML-KEM、ML-DSA 和 SLH-DSA 等新一代算法。 - [09:05](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=545s) 在企业级 AI / RAG 管道中，数据摄入阶段以及向量数据库与语义检索等环节汇集了大量高价值数据，是抗量子加固的优先防御重地。 - [11:13](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=673s) AI 安全必须贯穿微服务通信、模型权重保护和日志防泄漏，同时企业必须严格审查第三方供应商并推动其制定明确的 PQC 升级路线图。 - [12:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=727s) PQC 的核心定位是保护围绕 AI 运转的“信任基础设施”（如认证网关、传输链路与数字证书体系），而非直接加密神经网络模型参数本身。 - [13:28](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=808s) 企业应建立结构化评估矩阵，依据数据敏感度、机密性生命周期和架构复杂度，优先在隔离沙箱或实验环境中开展 PQC 试点验证。 - [16:42](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1002s) PQC 迁移绝非简单的“算法替换”，其安全性高度依赖高质量熵源与密钥管理，同时需防范 PQC 庞大密钥与签名体积引发的系统延迟与网络开销。 - [18:52](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1132s) 各层级团队应协同推进 90 天前瞻性落地规划，利用前 30、60、90 天窗口期完成加密资产清单（CBOM）盘点与风险测算。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=4s) “先窃取，后解密”：后量子时代企业 AI 数据面临的新型威胁 来自 Visa 的应用与后量子密码学资深网络安全工程师 Camille Nwaefulu 开场向与会者问好。她开玩笑说自己身处美东时间，整个人的状态还有点倒时差，但今天将深入探讨如何在企业级层面利用后量子算法保护人工智能（AI）数据。为了让各类背景的参会者都能理解，她将内容在技术细节与通俗概念之间进行了平衡。 随着企业加速拥抱 AI 技术，数据的流动性与复杂性大幅上升，敏感数据已经不再像过去那样安全。在当前的密码学安全讨论中，企业必须高度警惕一个核心概念——**“先窃取，后解密”（Harvest Now, Decrypt Later, 简称 HNDL）**。量子计算的发展正对现有的公钥密码学体系（Public Key Cryptography）构成深远的长期威胁。在现代企业复杂的 AI 架构中，涉及数据检索、向量数据库（Vector Databases）、API 调用、云基础设施、日志记录与监控等诸多环节，而所有这些环节的信任基石都建立在底层密码学的安全性之上。 现代组织在数据传输中普遍采用两套加密体系：一方面使用对称加密算法（如 AES-128、AES-256）保护数据内容，另一方面依赖 RSA、ECC（椭圆曲线密码学）等公钥密码体制来建立密钥分发与信任连接。然而，攻击者当前已经开始在网络传输中大肆拦截并保存加密流量。企业中的很多高敏感数据（例如秘密专利、核心商业机密等）当下看似没有立即被破解的风险，但黑客会将截获的密文保存 5 年、10 年甚至 15 年，直到未来足够强大的容错量子计算机（Fault-Tolerant Quantum Computers）问世，届时这些历史密文便会被瞬间破解并全面暴露。 --- ### [03:02](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=182s) 将量子风险转化为行动：企业应对的四大阶段框架 面对量子计算对现有密码学构成的系统性威胁，企业不能坐以待毙，而是需要将量子风险转化为清晰的执行路线图。演讲者将这一应对战略拆解为四个有序的实施阶段： ```mermaid flowchart LR accTitle: Quantum Risk Action Framework subgraph S1["阶段一：识别风险"] A["执行量子威胁分析<br/>排查潜在薄弱环节"] end subgraph S2["阶段二：确定优先级"] B["高风险数据分类分级<br/>合理分配预算与人力"] end subgraph S3["阶段三：建立就绪能力"] C["构建敏捷密码架构<br/>验证软硬件与HSM兼容性"] end subgraph S4["阶段四：推进迁移落地"] D["替换易受攻击公钥体系<br/>跨部门协同落地PQC标准"] end A --> B B --> C C --> D ``` - **第一阶段：识别风险（See the Risk）**：开展全面的量子威胁分析，全面摸排并识别企业架构中所有易受量子攻击影响的潜在脆弱点。 - **第二阶段：确定优先级（Set the Priority）**：企业并非所有数据都需要耗费巨资进行顶级防护，因此必须识别出真正的高风险资产。根据数据分类分级（如公开信息无需过度投入，但商业机密、医疗健康数据等核心资产必须严密保护），将有限的资金和人力精准投放到最关键的防护节点上。 - **第三阶段：建立就绪能力（Build Readiness）**：构建**敏捷密码（Crypto-Agile）**架构。引入新算法和新型加密机制时，必须确保不会破坏现有的系统架构与业务稳定性。许多人常说“我想直接上线 PQC、部署这个部署那个”，但实际必须评估基础设施的就绪情况；例如企业在某一容量和配置下运行硬件安全模块（HSM）时，升级改造前必须全面验证软硬件各个组件之间的兼容性。 - **第四阶段：推进迁移（Make the Transition）**：使用 PQC 标准逐步替换掉依赖传统公钥机制的脆弱组件，并将具体运维任务下发给跨职能团队。后量子密码迁移绝不是单一技术团队的事，必须获得高管层（Executive Team）、技术研发团队、法务合规团队（Legal Teams）的通力协作与全方位支持，才能确保全公司标准统一推进。 --- ### [05:10](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=310s) 量子威胁的本质：公钥密码体系 vs 对称加密算法 在技术机理层面，演讲者详细辨析了公钥密码学与对称密码学在面对量子计算机冲击时的不同表现： ```mermaid flowchart TD accTitle: Cryptography Comparison Under Quantum Threat subgraph PK["公钥密码学 (RSA, ECC)"] PK1["依赖大整数分解与离散对数等数学难题"] PK2["易被大规模量子计算机完全攻破"] PK3["威胁影响：密钥交换、数字签名、身份证书、信任链"] end subgraph Sym["对称密码学 (AES-128, AES-256)"] Sym1["不依赖易被量子算法破解的传统数论难题"] Sym2["量子攻击无法像破解公钥那样直接击穿"] Sym3["强对称加密仍将长期作为 PQC 架构核心组成部分"] end ``` 企业当前面临的当务之急，是将脆弱的公钥密码体系迁移至抗量子的替代方案。RSA 和 ECC 等经典公钥算法依赖于特定数学难题（如大整数分解与离散对数问题），而大规模量子计算机能够在多项式时间内轻松解出这些数学题，进而彻底瓦解现有的密钥协商、数字签名、数字身份证书以及网络信任链条。相比之下，以 AES 为代表的强对称加密算法不会以相同的方式被彻底破解，因此在后量子安全架构中，强对称加密依然会扮演极其重要的底层角色。 --- ### [06:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=367s) HNDL 威胁规划：数据保密寿命与迁移周期的博弈 针对“先窃取，后解密”（HNDL）的风险模型，企业在制定应对策略时必须建立科学的规划方程，综合考量**数据保密寿命（Confidentiality Lifetime）**与**系统迁移周期（Migration Timeline）**的重叠关系。 如果某类数据的价值仅维持在未来一到四年内，那么其面临的量子威胁相对有限，企业不必为此投入过度的改造资金或人力资源，从而避免资源浪费；但对于生命周期长达数十年、具有长期战略价值的资产，黑客截获并无限期存储密文的行为就构成了致命威胁。企业必须精准定位需要长期保密的资产，防止通信流量在网络中被黑客静默记录、拦截并留待日后解密。 --- ### [07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) 什么是后量子密码学（PQC）与 NIST 新一代算法标准 演讲者特别澄清了一个广泛存在的认知误区：**“后量子密码学（PQC）本身并不需要量子计算机来运行。”** PQC 指的是能够抵御经典计算机和未来量子计算机双重攻击的全新加密算法。这些算法完全可以运行在现代企业现有的笔记本电脑、通用服务器、云环境以及常规企业基础设施之上。 PQC 与经典公钥算法的核心差异在于底层的数学构造：它不再依赖大整数因子分解等易被量子计算快速求解的数论难题，而是基于目前被数学界公认为极其困难、即便借助量子算力也难以攻破的新型数学问题（如格密码等）。 目前，美国国家标准与技术研究院（NIST）已经正式发布并标准化了新一代密码工具集： - **ML-KEM（基于模块格的密钥封装机制）**：原名 **Kyber**，用于实现安全握手与密钥交换。 - **ML-DSA（基于模块格的数字签名算法）**：原名 **Dilithium**，充当数字印章（Digital Seal）以验证完整性与身份。 - **SLH-DSA（无状态无哈希数字签名算法）**：原名为 SPHINCS+ 的哈希基签名替代方案。 --- ### [09:05](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=545s) 企业 AI 架构全流程中的信任链与数据暴露面 为了让企业清晰了解在哪些环节需要部署 PQC 防御，演讲者拆解了一个典型的企业级 AI / RAG 系统的核心工作流与数据暴露面： ```mermaid flowchart TD accTitle: Enterprise AI Data Flow and Security Boundaries U["用户发起 Prompt"] --> ID["身份管理 (Identity Management)<br/>校验用户权限与访问控制"] ID --> GW["AI 网关 (AI Gateway)<br/>速率限制、安全协议管控"] GW --> DS["企业数据源 (Enterprise Data Sources)<br/>业务文档与知识库底层"] GW --> VDB["向量数据库 (Vector Database)<br/>语义检索与实时上下文增强"] VDB --> FM["基础模型 (Foundational Model)<br/>信息推理与响应生成"] FM --> ML["监控与日志 (Monitoring & Logs)<br/>捕获行为、审计与备份"] classDef secure fill:#e8f4f8,stroke:#2b7cff,stroke-width:2px; class ID,GW,DS,VDB,FM,ML secure; ``` 在整个企业 AI 处理管道中，每一个节点都深度依赖底层密码学建立的安全信任： 1. **身份管理（Identity Management）**：在交互发生前，验证用户访问权限并进行身份鉴权。 2. **AI 网关（AI Gateway）**：充当守门人角色，执行限流、安全策略过滤与通信协议管控。 3. **企业数据源（Enterprise Data Sources）**：承载核心业务文档与信息的底层存储。 4. **向量数据库（Vector Database）**：提供语义检索（Semantic Retrieval）与实时上下文注入，消除大模型幻觉并支撑精准问答。 5. **基础模型（Foundational Model）与监控日志（Monitoring and Logs）**：模型生成响应后，所有操作活动都会被记录进日志系统，用于合规审计、系统备份与未来性能优化的反馈闭环。 整套工作流中，TLS 协议负责保障各组件之间的通信加密，身份认证系统负责校验访问实体。然而，在**数据摄入（Data Ingestion）阶段**——即原始数据输入系统时的传输与接入点，以及**向量存储与语义搜索（Vector Stores & Semantic Search）**等环节，大量高价值数据极易处于暴露状态，成为后量子时代必须优先加固的防御重地。 ### [11:13](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=673s) 保护企业级 AI 数据流与第三方供应链安全 在构建安全的企业级 AI 系统时，单纯保护外围边界远远不够，安全防护必须贯穿数据流动的每一个具体环节。当系统执行向量嵌入与检索时，数据面临着在服务之间流转的暴露风险，企业必须全面强化内部微服务与 API 的通信安全。紧接着是模型处理环节，推理引擎的运行环境和核心模型权重必须得到严密防护；而日志记录与可观测性系统同样不容忽视，必须杜绝任何敏感数据通过调试信息或系统日志外泄的可能。此外，针对数据保留与备份体系，企业必须建立稳健的失效回滚机制，确保系统在遭遇意外时具备弹性恢复能力。 除了内部架构，第三方服务已经成为现代企业 AI 生态中最关键的潜在薄弱点之一。大量企业在构建 AI 管道时高度依赖外部 SaaS、第三方模型 API 或数据供应商，因此主动对接外部合作伙伴显得尤为紧迫。企业必须严格审查第三方是否采用了正确的加密方案以及合规的抗量子密码（PQC）标准；如果供应商尚未跟进，企业就必须推动其制定明确的技术路线图或推进计划，以确保整条供应链协同迈向抗量子安全标准。 ```mermaid flowchart LR accTitle: 企业级 AI 数据流关键防护节点 subgraph Core["核心计算与通信"] A["微服务 API 通信"] --> B["推理引擎与权重"] B --> C["向量嵌入与检索"] end subgraph Ops["运维与数据管理"] C --> D["日志与可观测性"] D --> E["数据备份与回滚"] end subgraph Supply["外部依赖"] E --> F["第三方服务与供应商"] end ``` - **微服务安全加固**：全面强化内部 API，防止向量检索与数据流转过程中的横向移动攻击。 - **推理与权重保护**：确保模型推理引擎在受信任环境中执行，严防核心模型权重资产被逆向或盗取。 - **日志防泄漏机制**：对观测日志进行脱敏与合规审计，消除敏感信息伴随系统运行泄漏的隐患。 - **第三方准入与路线图**：建立第三方供应商加密与 PQC 评估机制，对不达标的供应商设定明确的升级时间表。 --- ### [12:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=727s) PQC 在 AI 堆栈中的定位：保护信任基础设施而非模型本身 在梳理完 AI 数据流之后，核心问题随之而来：抗量子密码究竟应该部署在 AI 架构的哪一个层级？当前业内存在一个相当普遍的认知误区，即“人们往往误以为抗量子密码是用来直接保护 AI 模型本身的”。这并非事实真相。PQC 真正要捍卫的，是围绕人工智能运转的整套“信任基础设施”（Trust Infrastructure）。AI 系统并非孤立存在，从员工登录 AI 助手时的身份认证，到底层微服务之间密集的 API 调用与数据加密传输，整个技术栈都建立在加密信任链条之上。 具体而言，PQC 的防御覆盖范围可细化为多个核心层面：首先是量子抗性连接，专门用于防护 AI 网关、API 网关以及向量数据库的传输链路；其次是量子抗性的身份与信任体系，重点重构数字证书体系、工作负载身份验证以及安全的用户访问权限；随后是受信任的软件模型分发与签名验证；最后则是长期数据的抗量子存储保护，化解未来量子计算对当前截获数据的解密隐患。 - **破除认知误区**：PQC 的首要使命不是为神经网络模型参数加密，而是为保障 AI 正常运转的认证、传输与授权链条构建抗量子防线。 - **量子抗性传输连接**：重点覆盖企业 AI 网关、API 网关及向量数据库的外部与内部通信通道。 - **身份与凭证体系重构**：全面替换企业内部依赖传统公钥体系的数字证书与机器工作负载身份认证。 - **长期数据安全**：抵御“先截获、后解密”（Harvest Now, Decrypt Later）威胁，消除历史数据在量子时代的暴露风险。 --- ### [13:28](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=808s) 量子风险分级与实用迁移评估框架 并非所有 AI 工作负载所承受的量子风险都是等同的，信息安全的核心原则始终在于“优先级划分”，这一逻辑在 PQC 迁移过程中同样适用。诸如客户财务记录、医疗健康数据、商业秘密、源代码以及核心研发资料等高价值资产，往往受到严格的合规监管，需要在长达数年乃至数十年的时间跨度内保持绝对机密。相比之下，普通用户姓名等低风险字段或日常业务记录的保密周期要求则短得多；至于企业的市场营销文案、公开对外内容等公开数据，企业完全没有必要投入大量资金实施高强度的抗量子改造。 为了让迁移工作落地，企业需要引入一套结构化的评估与优先级排序框架： | 评估维度 | 核心考量要素 | 实施要点与关注指标 | |:--- |:--- |:--- | | **数据敏感度** | 数据分类等级划分 | 明确界定受限数据（Restricted）、受限 PII、机密数据与非机密数据 | | **机密性生命周期** | 数据留存时长评估 | 区分数据的安全有效周期（如短期留存 2 天、业务归档 5 年或法规要求 15 年） | | **密码资产暴露面** | 现有基础设施密码状况 | 摸清当前使用的加密算法类型、部署方式及其密钥强度，评估抗量子缺口 | | **迁移复杂度** | 架构重构难度与资源消耗 | 评估技术实现代价，防止在缺乏足够资源的情况下盲目切入复杂系统 | | **优先试点验证** | 隔离环境实测 | 率先在开发环境、沙箱或实验室内验证升级，确保平稳后再推向生产环境 | | **就绪度与互操作性** | 基础设施升级兼容性 | 监控新旧工具与系统的兼容性，保证各服务组件间的互操作性并持续归档风险 | --- ### [16:42](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1002s) 密码敏捷性、熵值基石与常见的迁移误区 企业在将 PQC 引入生产环境时，必须时刻秉持“密码敏捷性”（Crypto-Agility）原则，即加密算法的迭代升级绝不能破坏现有模型与业务系统的正常可用性。很多团队往往忽视了埋藏在 PQC 算法之下的信任根基——“熵”（Entropy）。抗量子密码的强度最终取决于高质量的随机性输入，缺乏可靠的随机数生成、薄弱的密钥管理、不可靠的运行环境以及粗糙的工程实现质量，都会使哪怕最先进的后量子算法彻底失效。完整的迁移应当遵循系统化的六步方法论：“发现（Discover）、分类（Classify）、定级（Prioritize）、设计（Design）、试点（Pilot）、扩展与治理（Scale and Govern）”。 在具体推进中，最致命的认知错误就是将 PQC 迁移当成简单的“算法替换”（Algorithm Swap）。抗量子密码不是即插即用的组件，如果脱离了对企业软硬件资产的全面盘点，盲目摸黑推进必将导致严重系统故障。更普遍的陷阱包括：只防御外围终端而忽视核心数据链路；轻信新算法具有“直接插入兼容性”，从而低估了后量子算法庞大的密钥与签名尺寸给系统延迟和计算吞吐带来的巨大负担；以及在孤立环境中闭门测试，忽视了与上下游真实生态的协同。 - **“非即插即用”警示**：“PQC 迁移绝不是一次简单的算法替换”，不能指望把新算法直接塞进现有系统就能平稳运行。 - **熵与四大信任基石**： 1. 高质量的随机性（Randomness） 2. 健壮的密钥管理（Key Management） 3. 安全可控的操作环境（Operational Environments） 4. 严谨的工程实现质量（Implementation Quality） - **性能与开销挑战**：PQC 算法的公钥与签名体积远大于传统 RSA 或 ECC，极易引发网络丢包、传输延迟激增或缓冲区溢出。 - **资产透明度优先**：必须在看清全套加密清单的前提下启动项目，严禁在缺乏资产全貌认知时实施盲目改造。 --- ### [18:52](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1132s) 组织角色分工与 90 天前瞻性落地规划 在抗量子演进的全局棋盘中，不同生态位角色的职责边界截然不同。对于严重依赖外部生态的小型企业或业务负责人而言，核心任务是看清并追问第三方的 PQC 技术路线图。小型企业自身缺乏底层研发资源，因此必须搞清数据在第三方平台上的保护细节：“数据究竟是如何存储的？密钥是如何受控的？是否使用了硬件安全模块（HSM）？采用的是服务端加密还是客户端加密？”小企业必须对这些链条建立起透彻认知。而对于大型企业管理者，则具备自主实施高规格 PQC 迁移的资源与能力，应当主导建立跨部门的治理机制，将技术开发、业务管理、文档编写、测试与合规团队紧密联合，消除跨部门信息孤岛。 对于一线技术团队，当务之急是立即启动全量资产盘点，全面审查内部所有的数据存储库与密码调用点。在此基础上，企业可以推行一个渐进式的 90 天准备框架。尽管全系统落地 PQC 必然远超 90 天，但将前 90 天拆解为 30 天、60 天与 90 天的明确阶段性目标，能够帮助团队在投入大规模技术改造之前，扎实完成战略规划、资源整合与风险测算，为企业在后量子时代保护 AI 数据筑牢防线。 - **小型企业负责人职责**：深入穿透第三方供应链，严格追验加密托管模式、HSM 硬件根证书保障及客户端加密支持情况。 - **大型企业高管职责**：打破部门壁垒，推动横跨架构、工程、法务合规及安全测试的协同治理机制。 - **技术工程师行动准则**：全面审计加密库调用、证书有效期限及后端数据仓库，建立动态加密资产清单（CBOM）。 - **90 天准备期节奏**：利用前 30、60、90 天窗口期完成摸底、规划与资源准备，为后续长期的渐进式上线奠定制度与架构基础。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=4s) 先窃取，后解密 — Camille Nwaefulu > [03:02](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=182s) 我想直接上线 PQC、部署这个部署那个 — Camille Nwaefulu > [07:17](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=437s) 后量子密码学（PQC）本身并不需要量子计算机来运行。 — Camille Nwaefulu > [12:07](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=727s) 人们往往误以为抗量子密码是用来直接保护 AI 模型本身的 — Camille Nwaefulu > [16:42](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1002s) PQC 迁移绝不是一次简单的算法替换 — Camille Nwaefulu > [18:52](https://www.youtube.com/watch?v=-cZ1lkUJL6Y&t=1132s) 数据究竟是如何存储的？密钥是如何受控的？是否使用了硬件安全模块（HSM）？采用的是服务端加密还是客户端加密？ — Camille Nwaefulu ## Mentioned Visa Inc., Camille Nwaefulu, Ai4, RSA, ECC, AES-128, AES-256, NIST, ML-KEM, Kyber, ML-DSA, Dilithium, SLH-DSA, SPHINCS+, HSM ## Video description Camille Nwaefulu - Senior Engineer, Applied and Post-Quantum Cryptography, Visa Inc. As enterprises accelerate AI adoption, they are collecting, combining, and moving greater volumes of sensitive data across increasingly complex environments. This talk will examine post-quantum cryptography through a data-protection lens, exploring why organizations must consider not only how data is secured today, but how long that protection must remain effective. I will discuss the practical challenges enterprises face, including identifying high-value data, understanding where vulnerable cryptography is used, managing legacy systems, and evaluating vendor readiness. Attendees will learn how data classification, risk-based prioritization, small pilot initiatives, and crypto-agile design can support a realistic post-quantum roadmap. This session will help practitioners connect AI growth, long-term data exposure, and post-quantum preparation without requiring a deep background in quantum computing. [Source](https://www.youtube.com/watch?v=-cZ1lkUJL6Y)

Source quotes

[0:04] Camille Nwaefulu: 先窃取，后解密 [3:02] Camille Nwaefulu: 我想直接上线 PQC、部署这个部署那个 [7:17] Camille Nwaefulu: 后量子密码学（PQC）本身并不需要量子计算机来运行。 [12:07] Camille Nwaefulu: 人们往往误以为抗量子密码是用来直接保护 AI 模型本身的 [16:42] Camille Nwaefulu: PQC 迁移绝不是一次简单的算法替换 [18:52] Camille Nwaefulu: 数据究竟是如何存储的？密钥是如何受控的？是否使用了硬件安全模块（HSM）？采用的是服务端加密还是客户端加密？

Source quotes

[7:17] Camille Nwaefulu: Post-quantum cryptography does not require quantum computers. [12:07] Camille Nwaefulu: People believe that post-quantum cryptography is going to protect the AI model itself. And that is not exactly the correct case. Instead, it's going to protect the trust infrastructure around the actual artificial intelligence. [18:34] Camille Nwaefulu: do not interrupt your models and your infrastructure. [18:34] Camille Nwaefulu: get further together.

Original transcript

[0:04] Hi, good afternoon, good morning. I am in  Eastern time, so I'm all over the place. [0:09] Um, so good morning. I'm Camille Weffelman. I am a  senior cybersecurity engineer. I focus in applied [0:15] and post-quantum cryptography. So, today we are  going to talk about enterprise level protection [0:21] of AI data utilizing post-quantum algorithms.  So, I'm trying to make it a mix of non-technical [0:28] and technical so that everybody is able to  conceptualize what we're being um discussed here. [0:33] Um, so right now we're going to talk  about protecting enterprise AI data in [0:37] the post-quantum era. So, currently right now  with the increase of artificial intelligence, [0:42] sensitive data is not as secure as  it may have been once upon a time. [0:46] Um, with that being spoken about, we have a  concept called harvest now, decrypt later, [0:51] which is really, really something that enterprises  need to start focusing on and preparing for. [0:58] So, the threat here is going to be that  these advances in quantum cryptography are [1:02] posing a long-term risk to the current public key  cryptographic standards. Um, so right now we need [1:07] to make sure that enterprises are prepared because  there are going to be different data retrievals, [1:13] vector databases, APIs, cloud infrastructures, and  different logging, monitoring, so many different [1:20] instances where we're going to need to make  sure that trust is built on secure cryptography. [1:27] So, while talking about harvest now, decrypt  later, we're focusing on what's happening today. [1:33] So, today we have our data transmission. With  that being said, organizations need to protect [1:39] this using Well, sorry. Organizations are  protecting this data using synthe- um uh [1:45] Um, symmetric encryption such as AES 256, um  128, etc., and public key mechanisms such as RSA [1:52] and ECC as they establish key trust. [snorts] In  interim, collection and storage, ad- adversaries [2:00] are able to intercept and store encrypted data  for future accesses. So, essentially now, if you [2:05] have an enterprise that you have sensitive data  in your organization, the the data may not seem [2:11] pressing right now, but there's going to be bad  actors who are collecting this data and basically [2:16] preserving it for 5, 10, 15 years until there's  strong enough cryptography to basically extract [2:22] that information and then makes it vulnerable. So,  if you have a secret patents or something that you [2:27] have currently, harvest now, decrypt later will  impact that if it is not protected adequately. [2:35] And then in the future, like I said,  fault-tolerant quantum cryptography, [2:40] um computers will emerge and basically  make the data vulnerable. [snorts] [2:45] So, basically, if you think about it, when you  have data going through two data centers or data [2:50] that's being protected over time, you really want  to make sure that they are using the strongest [2:55] form of protection possible. Um okay, so turning  quantum risk into action. So, the first portion [3:02] I have broken it down into four phases. So,  we're going to have our phase one, which is [3:07] you're going to see the risk. That is basically  going to be you need to execute a quantum threat [3:12] analysis and be able to identify the areas in  your enterprise that are potentially vulnerable. [3:19] Phase two is going to set the priority. You need  to identify the high-risk data that you have in [3:23] your organization. Not all the data that you're  currently protecting is going to be um necessary [3:29] to spend a lot of money on protecting. So, you  need to identify where you want to prioritize [3:34] your funding, where you want to prioritize your  manpower to ensure that the enterprise data [3:38] is adequately being protected according to  its classification levels. So, of course, [3:43] you're going to have data that might be public  information that doesn't need to be protected as [3:47] much as maybe your trade secrets or health data  or certain types of information of that sort. [3:53] The next is going to be builds readiness. So,  you're going to need to develop crypto agile [3:59] architectures, and basically make sure that  when you're entering this data into Sorry, [4:04] when you're entering these new forms of encryption  and um the algorithms that we're utilizing, [4:10] that it's not breaking your architecture, right?  It's one thing to implement something new without [4:15] actually causing your setup to fail. Okay? So,  a lot of times people say, "Oh, I want to get [4:21] into a PQC. I want to implement this. I want  to implement that." But, you need to make sure [4:25] that your infrastructure is prepared to do so. For  example, if you have HSMs running in one capacity, [4:31] and then you know that you want to maybe upgrade  your systems, you need to make sure that there is [4:35] compatibility amongst both as aspects. And then  the last part's going to be making the transition [4:41] the transition. You want to replace the quantum  vulnerable public key dependencies with the PQC [4:47] standards and assign the correct operational tasks  to cross-functional teams. So, the post-quantum [4:54] approach is going to be needed to be done amongst  different teams. You're going to have to have your [4:59] executive team on board, your technical teams on  board, your legal teams. Everybody needs to be on [5:04] board whenever you are implementing this so that  it can be consistent across the the board. So, the [5:10] quantum threat explained, right? So, the immediate  priority is going to be migration of the public [5:14] key systems to quantum-resistant alternatives.  So, we have on one side we have our public key, [5:20] it's going to be your RSA, your ECC, etc. And  then you're going to have your symmetric keys [5:24] on the side, um your AES. So, um vulnerable So,  we're looking [clears throat] on the left side, [5:31] you'll see that they're vulnerable to sufficient  crypto capable crypto quantum computers. [5:37] Primarily impacts the key establishment,  digital signatures, identity certificates, [5:42] and trust. Okay? And then on the other side,  we want to make sure that they're not broken [5:47] in the same manner as public key cryptography  and strong symmetric encryption remains part of [5:52] the post-quantum architecture. Okay, so today  with that being said that these algorithms [5:59] rely on mathematical problems that large-scale  quantum computers are expected to solve. Okay, so [6:07] So the quantum threat harvest now decrypt later  as we spoke about a little bit earlier before. [6:12] The main things are going to make sure that you're  able to capture, preserve and then decrypt later. [6:18] That is the current concern. So what we need to  focus on is making sure that the planning equation [6:24] in incorporates confidentiality lifetime and then  migration timelines over time to quantum risk. [6:33] So depending on how long you actually need  your data to be stored is going to matter. [6:38] If there's something that is not going to be  risky in the next year, two, three, four years, [6:44] you may not need to protect that is adequately  because again when we're implementing these things [6:48] into organizations, we want to make sure that  we're not costing unnecessary funding. We're not [6:53] misallocating our manpower. We want to make sure  that we are using it for the data that needs to be [6:59] protected. So you want to want to make sure that  you're going to record and intercept encrypted [7:05] traffic across the network, right? Then preserving  storing the intercepted ciphertext indefinitely [7:10] for future access. We don't want these things  to happen and then again decrypting later. [7:17] So when we're looking at what PQC is, you know,  post-quantum cryptography is essentially just [7:22] referring to these algorithms that are designed  to resist attacks from both classical computers [7:27] as well as future quantum computers. Okay, one  of the most important things that I do want to [7:32] emphasize is that post-quantum cryptography does  not require quantum quantum computers. Okay, [7:39] these algorithms run on the same laptops, the  same servers, the same cloud environments and [7:44] enterprise infrastructures that we already use  today. But, the difference is the mathematics. [7:49] Instead of relying on the mathematical problems,  such as integer factorizations, um these new ones [7:56] are essentially going to rely more so on um making  sure that we're utilizing the Sorry, I just got [8:03] confused. Utilizing these ones that don't rely  on mathematical problems, but instead of problems [8:09] that we believe to remain difficult. We want to  make sure that our encryption is difficult and [8:13] it's hard to crack, right? That's what's making  it very important, what's very needed is to make [8:19] sure that when we're upgrading to these new um  algorithms, that they are still strong. Okay, [8:26] so NIST has a standardized tool for establishing  secrets and trust, right? So, currently we have [8:31] these three very You guys should all be familiar  with these ones. We're going to have ML-KEM, [8:36] which is going to be the secure handshake, okay?  Then, we're going to have ML-DLS, I mean DSA, [8:42] and that's going to be our digital seal, and  then we have our SLH-DSA, which is our hash-based [8:47] alternative. So, I'm not going to go too deep into  these right now, but these are the You may also [8:52] know them as Kyber and Dilithium, but they have  been renamed as um K K ML-K KEM and then ML-DSA. [9:05] So, one question that, you know, I think that's  really important to know is kind of how to start, [9:11] what needs to be protected. So, we have the  identity management aspect, okay? So, basically [9:18] this is going to validate our user accesses  and permissions before the interaction start. [9:24] Then, we're going to have our AI gateway. So,  this is going to act as the gatekeeper for real [9:29] um rate limiting, security protocols, etc. Next,  [snorts] we're going to have our enterprise data [9:35] sources, and that's going to be the foundational  layer of our business information and documents. [9:42] Then we're going to have our  vector database, which is going [9:44] to basically provide the semantic retrieval  and real-time context to ground the response. [9:51] From there, we have our foundational model  as well as our monitoring and logs. So, [9:57] basically how it's going to be, you know, a user  will submit a prompt, the prompt is authenticated, [10:02] it passes through our gateways and our APIs, it  may retrieve information with vector databases [10:08] or enterprise [clears throat] knowledge bases,  and then these models process that information. [10:13] Responses are then going to be  generated, and then the logs are created, [10:17] which then goes into monitoring systems capturing  activity. Um backups are going to be stored, [10:23] feedback loops are going to be improved, and  future performance. So, basically, we just want to [10:30] make sure that every one of those systems depend  on are built on trust. So, of course, you know, [10:36] TLS is going to protect the communication,  and then identity systems verify the user, so. [10:46] So, now, where the data is exposed. So, data's  going to be exposed in these various areas. So, [10:53] one of the big ones is going to be in the Oh,  let me go back. So sorry. In the data ingestion. [10:58] Let me my layer right here. So, one of the  biggest areas is going to be data ingestion. [11:03] And essentially here, you want to make sure that  you're securing the raw inputs when they enter the [11:08] system. From there, we're looking at protecting  vector stores and semantic search processes, [11:13] which is in your embedding and your retrievals.  From the service-to-service communication, [11:19] we want to make sure that we're hardening any  of the internal APIs in the microservices. [11:24] Then [snorts] we're going to have our model  processing, which we want to make sure that [11:27] we're securing the inference um engines and  model weights. Then logging and observability, [11:33] we want to preserve prevent any source of data  leaks and then we're going to have our retention [11:37] and backups where essentially you want to make  sure that we have fail rollbacks if necessary [11:43] and then third-party services. So, one of the  really big important things that enterprises [11:47] need to do is connect with their third-party  services and ensure that they are utilizing [11:52] the correct forms of encryption, the correct  forms of PQC and if they aren't then have a [11:57] pipeline or a program or a roadmap that's going  to help your organization get to it. [snorts] [12:07] So, where PQC is going to protect the enterprise  AI stack? Okay, so we've already identified where [12:13] the AI data flows are through the enterprise.  The next question becomes where does post-quantum [12:18] cryptography actually fit? And one of the biggest  misconceptions that I hear is that, you know, [12:24] people believe that post-quantum cryptography  is going to protect the AI model itself. [12:30] And that is not exactly the correct case. Instead,  it's going to protect the trust infrastructure [12:35] around the actual artificial intelligence.  So, when you're thinking about every secure [12:40] connection that makes an AI possible, we're  thinking about how employees are logging into [12:44] their AI assistants. We are looking at how their  identity needs to be authenticated. We're looking [12:50] at how APIs communicate with one another and those  communications also need to be encrypted. Um Okay, [12:57] looking over here we have our quantum resistant  connections and again that's protecting our API [13:01] gateway, our AI gateways, our vectors. Then we  have our quantum resistant identity and trust. [13:07] That's going to focus around our certificates,  our secure user workloads and identities. Then [13:14] we're going to have our trusted software models  and then we're also going to have our long-term [13:19] data protection. So, again making sure that you  are protecting your data from the long-term risks. [13:28] So, not every AI workload has the same quantum  risk, and people need to also prepare to make sure [13:35] the security all is all about prioritization.  The same principle is going to apply to [13:40] the post-quantum migration. So, for example,  customer financial records, health data, [13:46] um trade secrets, source codes, resource data,  those type of things are essentially going to need [13:52] to remain confidential for an extended amount of  years. As opposed to certain information that may [13:58] be lower level, maybe something like name, right?  We want to make sure that we're prioritizing the [14:04] data in the enterprise spaces to ensure that the  highest priorities and our most protected assets [14:10] are fully covered and they're protected. Um for  our medium priority sensitive business data, [14:17] of course, that's going to be a little bit lower  on the um rank, but essentially all of these are [14:23] still going to be very important. When we're  looking at our public data, that's marketing, [14:28] public-facing content, you won't want to invest  as much money protecting that information. So, [14:34] when we're looking at a practical prioritization  model, this is really what you want to pay [14:39] attention for. This is a This is a framework for  assessing that, right? So, we're going to have [14:43] our data sensitivity. Again, classification is  very important. You need to classify your data [14:47] before you can do anything. You need to understand  what kind of data you have, why you have the data, [14:53] you need to be able to classify it. Is it  going to be restricted? Is it going to be [14:57] restricted PI? Is it going to be confidential?  Is it going to be non-confidential? We need to [15:01] understand the data sensitivity. Then we need to  know the lifetime. How long is the data going to [15:06] be stored? Are we looking at it to be stored for  15 years? For 5 years? For 10 years? For 2 days? [15:13] Those type of things are going to matter whenever  we're determining the confidentiality lifetime. [15:18] Then the cryptographic exposure, understanding  currently where your infrastructure stands, [15:23] and then also making sure that you're  able to get it to the current and future [15:27] decryption protection era. So, currently  right now, you want to know what kind of [15:33] encryption your organization is using, how  they're using it, what type you're using, [15:39] and the strengths of those type of things. The  next thing is going to be my migration complexity. [15:44] Sometimes things might be a little bit too complex  to migrate, and you want to make sure that you're [15:49] not basically answering into a situation that  you can get out of, right? So, you want to [15:55] make sure that there's enough technical effort  and resource requirements for that transition. [16:00] The next one is going to be to migrate first,  right? So, what are you actually migrating first [16:06] and why are you migrating it? You want to make  sure that we are utilizing our dev environments, [16:11] our labs, our sandboxes to make sure before we  introduce this into our production environments, [16:16] that we have successfully deployed  it out in our test environments. [16:20] We're going to monitor readiness. We want to  make sure that all of the infrastructure upgrades [16:25] are viable, right? We don't want to make We don't  want situations where there is incompatibility. We [16:31] want to make sure that there is interoperability  amongst the different tools and the use cases. [16:36] And then you want to make sure  that you're documenting your risk. [16:42] So, crypto agility, right? Again, kind of going  back into what I just stated, you want to make [16:47] sure that whenever you're implementing and  you're integrating cryptography and PQC [16:51] into your infrastructure, that you're making  sure that it's not going to break your model. [16:57] So, this is just for viewing  purposes for the sake of time. [17:05] The next thing that's really  important is going to be entropy. [17:09] So, that's going to be the trust foundations  beneath the PQC. And the four components [17:13] that really go into that that going to be um  randomness, then key management, and then operate [17:19] operational environments and implementation  quality. So again, you want to make sure that [17:27] you have all of these components so that you're  able to have that stress foundation successfully. [17:35] And here is one more framework where basically you  can utilize this in your own enterprise on how to [17:42] um migrate. So again, you want to discover,  you want to classify, prioritize, design, [17:48] pilot, and then scale and govern. [17:54] So PQC migration is not an algorithm swap,  okay? So essentially, you can't just drop [18:00] it in and think that it's going to work.  You want to make sure that you're starting [18:04] with visibility. So if you start without  visibility, you're going to end up having [18:09] failures. You need to have a good understanding  of your inventory. You need to fully understand [18:14] everything that you have in your enterprise so  that you can successfully identify where to start. [18:20] The next is going to be making the mistake  of only protecting your endpoints, okay? [18:26] The next mistake would be assuming  drop-in compatibility. You want to [18:30] make sure that the performance impacts and  complexities of the larger key and signature [18:34] sizes do not interrupt your models and your um  infrastructure. Then ignoring trust foundations, [18:42] testing um in isolation, and then  lastly, migrating without crypto agility. [18:52] So your role in the post-quantum transition.  So if you're a business owner, um you need to [18:57] make sure that you are very familiar and aware of  the type of PQC roadmaps that your third parties [19:03] are utilizing. A lot of small businesses are  using a significant amount of third parties, [19:07] so you need to understand how is your data being  protected? Um how is the keys being stored? Are [19:12] you guys utilizing HSMs? Are you got hardware  security modules? Are you guys using, you know, [19:18] server-side encryption or client-side encryption?  Like you need to have a very robust understanding [19:23] of how the third parties are protecting your  data because you don't have access to it because [19:29] you're a smaller organization. But if you're  an enterprise leader, you have the ability to [19:35] implement serious levels of post-quantum  cryptography. So you need to make sure that [19:40] you're establishing governance policies, you're  implementing the correct teams. You need to make [19:45] sure that you are um working amongst the technical  teams, the managers, the non-technical teams, [19:51] the documentation teams, the testing, everybody  so that you guys can be on the same page about [19:56] integrating it because integrating post-quantum is  not an easy thing sometimes, but when everybody's [20:01] on the same page, you're able to get further  together. And of course the technical teams, [20:06] make sure that you are um taking inventory, you're  looking at all of your data stores for success. [20:15] And so yeah, this is just like a 90-day  implementation plan that you can take through. [20:20] And of course it will take probably longer  than 90 days to implement um PQC into your [20:25] infrastructure. However, this is a good  opportunity for you to use the 30 days, [20:29] the 60 days, and 90 days to prepare, plan, and  put together resources before you fully integrate. [20:36] And this is just again for additional  reading, strategic pillars that you're [20:40] able to implement in your organization  and kind of go deeper into. And I'll [20:45] leave that there for you guys to finish your  pictures. And that is it. Thank you very much.