DHS Warns AI Convergence Accelerates Quantum Breach of US Encryption

Presentation by Garfield Jones, U.S. Department of Homeland Security — Ai4

Ai4

A bidirectional feedback loop between artificial intelligence and quantum computing is compressing timelines to break standard public-key cryptography, exposing intercepted data to adversaries before 2035 federal defense deadlines.

AI-assisted summary of Presentation by Garfield Jones, U.S. Department of Homeland Security — original source. Check the original for full context.

Source published:

Detailed notes

## Background This video was published on YouTube by Ai4 on August 28, 2025, featuring a presentation by Garfield Jones, Associate Chief of Strategic Technology at the U.S. Department of Homeland Security. The program description highlights the emerging, bidirectional research between artificial intelligence and quantum computing, noting that much of the existing discourse around large language models and quantum capabilities remains grounded in early hypotheses and ongoing research.

## TL;DR Garfield Jones outlines the technical imperatives required to unite artificial intelligence and quantum computing, emphasizing that today's LLMs require dedicated quantum algorithms, novel data representations, and mature fault-tolerant hardware to function on quantum systems. He warns of the severe cybersecurity risks posed by Cryptanalytically Relevant Quantum Computers (CRQCs), which threaten public-key infrastructure via Shor's algorithm and enable adversarial 'Harvest Now, Decrypt Later' campaigns. While quantum computing can solve complex LLM reasoning bottlenecks and generate high-fidelity synthetic data, it simultaneously empowers malicious actors to accelerate exploit generation and deanonymize private datasets. Consequently, Jones underscores the critical importance of transitioning to NIST post-quantum cryptography standards to meet National Security Memorandum 10 (NSM-10) mandates ahead of the projected 2035 encryption horizon.

## Atmosphere & style The presentation maintains an authoritative, sober, and pragmatic tone typical of senior federal defense briefings. Jones balances genuine enthusiasm for quantum-assisted AI breakthroughs with a serious, risk-conscious posture concerning national security vulnerabilities. The delivery feels calculated and informative, shifting seamlessly from theoretical physics and algorithmic mechanics to pressing regulatory timelines and strategic deterrence.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) Merging artificial intelligence with quantum computing requires three prerequisites: dedicated quantum LLM algorithms, novel quantum data representations, and mature quantum hardware execution. - [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) The advent of a Cryptanalytically Relevant Quantum Computer (CRQC) running Shor's algorithm threatens to completely compromise asymmetric public-key cryptography, including RSA and ECC. - [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) NIST has released the first post-quantum cryptography (PQC) standards (FIPS 203, 204, and 205) to safeguard digital infrastructure against future quantum cryptanalysis. - [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) Quantum computing can dramatically accelerate LLM training cycles and enable high-fidelity synthetic data generation when real-world datasets are limited. - [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) Vulnerabilities in asymmetric key exchanges expose AI training pipelines to quantum-enabled man-in-the-middle attacks and data poisoning. - [11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) Quantum computing holds the potential to overcome structural bottlenecks in modern LLMs, particularly in complex logic and abstract reasoning. - [11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) Adversaries are actively executing 'Harvest Now, Decrypt Later' (HNDL) strategies, archiving encrypted communications today to decrypt them once CRQCs emerge. - [13:42](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=822s) The combination of quantum optimization and AI enables malicious actors to deanonymize private datasets and generate potent cyber exploits at unprecedented speeds. - [13:42](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=822s) A bidirectional acceleration loop exists where AI speeds up quantum hardware design and quantum computing enhances AI, compressing timelines toward CRQC feasibility. - [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) National Security Memorandum 10 (NSM-10) mandates post-quantum cryptography migration across federal systems and commercial defense contractors by 2035, with active targets aiming for 2030. - [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) Achieving fault-tolerant quantum computing requires tracking logical rather than raw physical qubits, operating at an approximate ratio of 1,000 physical qubits to one logical qubit.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) Introduction and the Emerging Intersection of AI and Quantum Computing

Opening his presentation at Ai4, Garfield Jones, Associate Chief of Strategic Technology at the U.S. Department of Homeland Security, reflects on the dramatic surge of interest in quantum technology. A few years ago, quantum was far from a popular or trendy subject—at his very first quantum computing conference appearance, only about ten people were in the audience. Today, the landscape has transformed dramatically, with major industry players like AWS, Microsoft, and IBM heavily pushing research releases and quantum milestones. While speaking in an official government capacity and strictly maintaining vendor neutrality, he emphasizes that the underlying research is accelerating quickly.

```mermaid flowchart TD accTitle: Three Key Developments for Quantum AI Integration subgraph Prerequisites["Prerequisites for Merging AI and Quantum"] A["1. Quantum-Tailored LLM Algorithms<br/>(Algorithmic shift from classical)"] B["2. Quantum-Suitable Data Representations<br/>(Redefining foundational AI data structures)"] C["3. Hardware Scalability & Capabilities<br/>(Sufficient QPU execution capacity)"] end Prerequisites --> D["Full Quantum-Accelerated AI / LLMs"] ```

However, seamlessly merging artificial intelligence with quantum computing is not simply a plug-and-play operation. Current large language models (LLMs) are architected exclusively around classical computing frameworks and will not run effectively on existing quantum systems. For quantum computing to genuinely enhance AI, three core technical developments must take place: - **Dedicated Quantum LLM Algorithms**: New algorithmic architectures designed specifically for the mechanics of quantum processors rather than classical binary logic. - **Novel Quantum Data Representations**: Fundamentally redesigned data structures capable of encoding information in a manner that quantum processors can natively ingest and manipulate. - **Maturing Quantum Hardware Execution**: Maturation of quantum hardware to a capability threshold where it can reliably execute large language model workloads at scale.

---

[03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) Quantum Computing Paradigms and the Cybersecurity Threat of CRQC

To contextualize the technology, quantum research and development splits broadly into two primary architectural avenues: special-purpose quantum computing and universal (general-purpose) quantum computing. Special-purpose systems focus on highly niche computational tasks, centered heavily on mathematical optimization and quantum annealing. While annealing remains one of the most widely adopted practical quantum computing approaches today, government security bodies like the Cybersecurity and Infrastructure Security Agency (CISA) are far more focused on universal, gate-based systems.

```mermaid flowchart LR accTitle: Quantum Computing Paradigms and Threat Vectors subgraph Approaches["Quantum R&D Approaches"] direction TB SP["Special-Purpose Quantum<br/>• Quantum Annealing<br/>• Specific Optimization"] UP["Universal / General Purpose<br/>• Gate-based execution<br/>• Full fault tolerance"] end

subgraph SecurityImpact["Cryptographic Impact"] CRQC["CRQC<br/>(Cryptanalytically Relevant<br/>Quantum Computer)"] Shor["Shor's Algorithm (1994)<br/>Breaks Asymmetric Encryption<br/>(RSA, ECC, PKI, Signatures)"] Grover["Grover's Algorithm<br/>Weakens Symmetric Encryption<br/>(Requires larger key sizes)"] end

UP --> CRQC CRQC --> Shor CRQC --> Grover ```

The primary concern for cybersecurity leaders is the eventual arrival of a **Cryptanalytically Relevant Quantum Computer (CRQC)**. Operating on general-purpose quantum architectures, a sufficiently powerful CRQC has the capability to dismantle the entire foundation of modern public-key infrastructure (PKI), rendering standard asymmetric encryption methods—including RSA, elliptic-curve cryptography (ECC), and digital signatures across emails and private messaging—entirely obsolete.

- **Physical and Hardware Constraints**: Realizing a CRQC requires exotic engineering breakthroughs, including specialized physical materials and extreme cryogenic cooling environments kept colder than deep space to maintain quantum coherence for long-form mathematical factoring. - **Shor's Algorithm**: Formulated by Peter Shor in 1994, this quantum algorithm provides an exponential speedup for prime integer factorization and discrete logarithms, exposing the vulnerability of current asymmetric schemes. - **NIST Post-Quantum Standards**: Acknowledging the threat in advance, NIST standardized the first quantum-resistant post-quantum cryptography (PQC) standards: **FIPS 203**, **FIPS 204**, and **FIPS 205**. - **Hybrid Computing Models**: Beyond pure paradigms, hybrid approaches—such as Quantum Approximate Optimization Algorithms (QAOA)—dynamically offload portions of computation to quantum processors before feeding results back to classical machines, optimizing workloads that classical computers still handle more efficiently.

---

[08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) Quantum Advantages, Synthetic Data, and Adversarial Threat Vectors in AI

The central problem of a CRQC lies in its exponential speed advantage over the best classical algorithms, allowing it to solve factoring problems in seconds that would otherwise take classical mainframes millennia. When applied constructively to artificial intelligence, this massive acceleration could dramatically slash the exorbitant training times currently required for foundational LLMs, while expanding the variety and depth of capabilities extractable from high-dimensional datasets.

| Dimension | Classical AI / Computing | Quantum-Assisted AI / Security | |:--- |:--- |:--- | | **Model Training Speed** | Bounded by classical GPUs/TPUs; long training cycles | Potential exponential acceleration via quantum routines | | **Data Requirements** | Requires massive scale real-world corpora | Efficient high-fidelity synthetic data generation with smaller samples | | **Asymmetric Cryptography** | Secure under classical factoring limits (RSA/ECC) | Fully compromised by CRQC running Shor's Algorithm | | **Symmetric Cryptography** | Standard AES architectures | Resilient against Shor's; targeted instead by Grover's Algorithm |

Quantum computing can also transform the generation of complex synthetic data, bridging gaps where real-world training examples are sparse or difficult to replicate. However, these same advancements introduce profound adversarial risks. Threat actors equipped with quantum capabilities could exploit asymmetric vulnerabilities to execute undetectable man-in-the-middle attacks, intercepting, poisoning, or manipulating AI training pipelines and access pathways before the data ever reaches the model. While symmetric databases remain comparatively safer against Shor's algorithm (requiring Grover's algorithm for quadratic speedup attacks instead), the asymmetric data-in-transit and key-exchange mechanisms securing AI infrastructure remain critically exposed.

[11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) Quantum Computing's Role in LLM Reasoning and Emerging Threat Vectors

While technological breakthroughs can emerge suddenly and catch organizations "behind the eight-ball," the current limitations of large language models—specifically in abstract reasoning and logic—represent a critical area where quantum computing (QC) can provide transformative capabilities. Today's LLMs face recognizable hiccups with complex logical processing, but the symbiotic fusion of AI and quantum computing introduces immense computing power that can bolster reasoning architectures. However, this convergence also introduces severe security challenges, most notably around training data manipulation.

A primary concern in this landscape is the "Harvest Now, Decrypt Later" (HNDL) strategy actively employed by adversaries. Hostile actors continuously siphon encrypted institutional and national data, warehousing it until a Cryptanalytically Relevant Quantum Computer (CRQC) becomes operational. Once adversaries achieve CRQC capability, they will be able to inspect the underlying data structures, identify systemic vulnerabilities, and manipulate training pipelines, making it imperative that defense and security stakeholders address post-quantum threats immediately.

- **Abstract Reasoning Enhancement**: Quantum computing offers the logical and computational leverage needed to solve foundational reasoning bottlenecks in modern LLMs. - **Harvest Now, Decrypt Later (HNDL)**: Adversaries intercept and store encrypted data today with the explicit goal of decrypting and analyzing it once CRQC systems are realized. - **Data Integrity Risks**: Decrypting legacy datasets exposes institutional data architectures to adversarial exploitation and targeted data manipulation.

---

[13:42](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=822s) Exploiting Algorithmic Weaknesses and Malicious QC-AI Capabilities

Quantum-enhanced LLMs present a stark asymmetrical advantage over standard classical LLMs. By leveraging quantum algorithms capable of solving complex optimization problems, adversaries can systematically expose hidden vulnerabilities within classical algorithms, defense frameworks, and machine learning models. Furthermore, bad actors can deploy quantum computing and AI concurrently to deanonymize private datasets and bypass conventional privacy safeguards.

Beyond surveillance and cryptanalysis, the union of quantum computing and generative AI significantly lowers the barrier to producing potent, highly tailored malicious content at unprecedented speeds. This bidirectional dynamic—where AI accelerates quantum computing breakthroughs and quantum computing turbocharges AI capabilities—drastically compresses the timeline toward achieving a fully functional CRQC.

```mermaid flowchart TD accTitle: Dual Convergence of AI and Quantum Computing

subgraph Convergence["Symbiotic Acceleration"] QC["Quantum Computing Capabilities"] <--> AI["Artificial Intelligence & LLMs"] QC --> CRQC["Accelerated Path to CRQC"] AI --> CRQC end

subgraph AdversarialThreats["Emerging Threat Vectors"] CRQC --> Opt["Algorithmic Optimization & Exploits"] CRQC --> Deanon["Deanonymization of Private Data"] CRQC --> MalGen["Rapid Malicious Content Generation"] end ```

- **Algorithmic Exploitation**: Quantum optimization methods can identify structural flaws across deployed classical security models. - **Data Deanonymization**: Quantum-assisted LLMs can reconstruct identities from anonymized datasets by correlating disparate, high-dimensional data points. - **Accelerated Weaponization**: Malicious synthetic content and cyber exploits can be generated much faster and with greater potency using quantum-boosted LLMs.

---

[15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) Timelines, Hardware Milestones, and Post-Quantum Cryptography Mandates

Examining the projected trajectory toward breaking standard public-key cryptography—including RSA and Elliptic Curve Cryptography (ECC)—reveals historical, pessimistic, and optimistic projection models. Under current optimistic forecasts, public-key encryption is slated to break around 2035, aligning directly with the deadlines established by National Security Memorandum 10 (NSM-10). NSM-10 mandates that government systems transition to Post-Quantum Cryptography (PQC) resistance, an obligation that directly extends to private-sector vendors doing business with the federal government, with internal implementation roadmaps pushing aggressively toward a 2030 target.

```mermaid flowchart LR accTitle: Quantum Hardware Progress Toward Fault Tolerance

subgraph Concepts["Theoretical Architectures"] M["Majorana"] O["Ocelot"] end

subgraph Physical["Demonstrated Hardware"] W["Google Willow"] end

subgraph Threshold["Fault-Tolerant Horizon"] LQ["1 Logical Qubit<br/>(~1,000 Physical Qubits)"] CRQC_Node["CRQC Readiness<br/>(NSM-10 / 2035 Target)"] end

Concepts --> Physical Physical --> LQ LQ --> CRQC_Node ```

Tracking hardware progress requires differentiating between raw physical qubits and fault-tolerant logical qubits, with roughly 1,000 physical qubits required to construct a single error-corrected logical qubit. While approaches like Majorana and Ocelot represent theoretical architectures, chips such as Google's Willow have demonstrated tangible milestones closer to the single logical qubit threshold. As AI is integrated into quantum design and error correction, transition timelines could compress rapidly, pulling the quantum cryptanalysis horizon well ahead of schedule:

- **NSM-10 Compliance**: Mandates federal agencies and commercial defense partners achieve full post-quantum cryptographic resistance by 2035, with current efforts striving toward 2030 readiness. - **Cryptographic Impact**: - **Shor's Algorithm**: Threatens asymmetric standards including RSA and Elliptic Curve Cryptography. - **Grover's Algorithm**: Targets symmetric ciphers such as AES, necessitating larger key sizes. - **Harrow-Hassidim-Lloyd (HHL) Algorithm**: Enables exponential speedups in solving linear systems and optimization equations. - **Qubit Scalability Metric**: Hardware evaluations must account for logical rather than raw physical qubits, operating at a ~1,000:1 physical-to-logical ratio. - **AI-Driven Acceleration**: Infusing AI into quantum engineering could accelerate quantum readiness lines, making proactive PQC migration vital before technological surprise occurs.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) Current large language models (LLMs) are architected exclusively around classical computing frameworks and will not run effectively on existing quantum systems. — Garfield Jones

> [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) A sufficiently powerful CRQC has the capability to dismantle the entire foundation of modern public-key infrastructure (PKI), rendering standard asymmetric encryption methods—including RSA, elliptic-curve cryptography (ECC), and digital signatures across emails and private messaging—entirely obsolete. — Garfield Jones

> [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) Threat actors equipped with quantum capabilities could exploit asymmetric vulnerabilities to execute undetectable man-in-the-middle attacks, intercepting, poisoning, or manipulating AI training pipelines and access pathways before the data ever reaches the model. — Garfield Jones

> [11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) Hostile actors continuously siphon encrypted institutional and national data, warehousing it until a Cryptanalytically Relevant Quantum Computer (CRQC) becomes operational. — Garfield Jones

> [13:42](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=822s) By leveraging quantum algorithms capable of solving complex optimization problems, adversaries can systematically expose hidden vulnerabilities within classical algorithms, defense frameworks, and machine learning models. — Garfield Jones

> [13:42](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=822s) This bidirectional dynamic—where AI accelerates quantum computing breakthroughs and quantum computing turbocharges AI capabilities—drastically compresses the timeline toward achieving a fully functional CRQC. — Garfield Jones

> [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) NSM-10 mandates that government systems transition to Post-Quantum Cryptography (PQC) resistance, an obligation that directly extends to private-sector vendors doing business with the federal government, with internal implementation roadmaps pushing aggressively toward a 2030 target. — Garfield Jones

> [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) Hardware evaluations must account for logical rather than raw physical qubits, operating at a ~1,000:1 physical-to-logical ratio. — Garfield Jones

## Mentioned Garfield Jones, U.S. Department of Homeland Security, Cybersecurity and Infrastructure Security Agency (CISA), Ai4, AWS, Microsoft, IBM, Google, Google Willow, National Institute of Standards and Technology (NIST), Peter Shor, National Security Memorandum 10 (NSM-10), FIPS 203, FIPS 204, FIPS 205, Shor's Algorithm, Grover's Algorithm, Harrow-Hassidim-Lloyd (HHL) Algorithm, Quantum Approximate Optimization Algorithm (QAOA), Majorana, Ocelot, RSA, Elliptic Curve Cryptography (ECC), AES

## Video description Garfield Jones, Associate Chief of Strategic Technology, U.S. Department of Homeland Security

Abstracts

There has been increasingly more research conducted on ways AI can help advance quantum computing (QC) and vice-versa. Most QC information regarding large language models (LLMs) is based on incomplete research or hypotheses of researchers.

[Source](https://www.youtube.com/watch?v=aVT9PhD88Pw)

中文简报

DHS警告：AI融合加速量子破解美国加密体系 人工智能与量子计算之间的双向反馈循环正在缩短破解标准公钥密码体系的时间表，使被截获的数据在2035年联邦防卫截止期限前面临对手威胁。 Garfield Jones概述了融合人工智能与量子计算所需的技术要件，强调当今的LLM需要专用的量子算法、新型数据表征以及成熟的容错硬件才能在量子系统上运行。他警告称，具备密码分析相关能力的量子计算机（CRQC）带来了严峻的网络安全风险，这些计算机可通过Shor算法威胁公钥基础设施，并助长对手开展“先窃取，后解密”（Harvest Now, Decrypt Later）活动。尽管量子计算能够解决复杂的LLM推理瓶颈并生成高保真合成数据，但它同时也能让恶意行为者加速漏洞利用生成并对私有数据集实施去匿名化。因此，Jones强调了向NIST后量子密码学标准过渡的关键重要性，以在预期的2035年加密时限之前满足第10号国家安全备忘录（NSM-10）的要求。 将人工智能与量子计算深度融合需满足三大核心前提：开发专用的量子大语言模型（LLM）算法、重构量子适用的底层数据表征，以及实现硬件执行能力的成熟扩展。 (0:04) 具备破解密码学能力的量子计算机（CRQC）通过 Shor 算法将彻底攻破 RSA、ECC 等现代非对称公钥加密体系，促使 NIST 推出首批后量子密码（PQC）标准（FIPS 203、204 和 205）。 (3:35) 量子计算可大幅缩短 LLM 的庞大训练周期并高效生成高保真合成数据，但非对称密钥交换环节的脆弱性也让 AI 训练流水线暴露于量子中间人攻击与数据投毒风险之中。 (8:03) 量子计算有望化解 LLM 在抽象逻辑推理上的结构性瓶颈，然而敌对势力正在执行“先收集、后解密”（HNDL）策略，并利用量子优化与 AI 协同加速敏感数据反匿名化及高危网络漏洞利用。 (11:30) 美国第 10 号国家安全备忘录（NSM-10）要求政府系统在 2035 年前全面实现抗量子迁移（力争 2030 年完成），当前硬件发展仍受制于约 1,000:1 的物理量子比特与容错逻辑量子比特转换比例门槛。 (15:30)

Reading summary

According to Garfield Jones, Associate Chief of Strategic Technology at the U.S. Department of Homeland Security, the convergence of artificial intelligence and quantum computing could resolve major computational and reasoning bottlenecks, but it simultaneously accelerates severe cybersecurity risks that threaten foundational encryption infrastructure. Technical Prerequisites for Quantum-Accelerated AI (0:04) Integrating artificial intelligence with quantum computing is not an immediate plug-and-play transition. Current large language models are engineered exclusively for classical computing frameworks and binary logic, rendering them incompatible with existing quantum systems. Jones identifies three fundamental hurdles that must be addressed: developing dedicated quantum LLM algorithms, establishing novel quantum data representations capable of native ingestion by quantum processors, and scaling quantum hardware to reliably support large-scale language model workloads. Once achieved, quantum integration could drastically reduce the massive training cycles currently required by foundational models. It also offers potential solutions to the abstract reasoning and logical processing constraints observed in modern LLMs, while enabling high-fidelity synthetic data generation when empirical training samples are scarce. Cryptographic Disruption and Post-Quantum Standards (3:35) The primary institutional concern surrounding quantum advancement centers on universal, gate-based systems maturing into a Cryptanalytically Relevant Quantum Computer (CRQC). Running Shor's 1994 algorithm, a functional CRQC would achieve an exponential speedup in solving prime factorization and discrete logarithms, effectively dismantling modern public-key infrastructure—including RSA, elliptic-curve cryptography, and standard digital signatures. Symmetric encryption like AES remains resilient against Shor's algorithm, though it faces quadratic speedup vulnerabilities via Grover's algorithm that necessitate larger key sizes. To counter this vulnerability, the National Institute of Standards and Technology (NIST) has released the first post-quantum cryptography standards: FIPS 203, FIPS 204, and FIPS 205. In the interim, hybrid frameworks like the Quantum Approximate Optimization Algorithm (QAOA) dynamically route select computational tasks to quantum processors while returning remaining workloads to classical systems. Adversarial Exploitation and Data Manipulation (8:03) The intersection of quantum computing and AI introduces dangerous threat vectors, most immediately through 'Harvest Now, Decrypt Later' (HNDL) campaigns. Hostile actors are actively stockpiling encrypted institutional and national data to decrypt once a CRQC is developed. Beyond retroactive exposure, decrypted legacy data allows adversaries to map structural vulnerabilities and tamper with machine learning data pipelines before training begins. Jones warns that combining quantum computing with generative AI provides asymmetric advantages to malicious actors. Quantum optimization techniques can uncover subtle flaws in deployed classical defense models, bypass privacy protections by correlating disparate datasets to deanonymize private records, and exponentially speed up the generation of targeted malicious content. Hardware Milestones and NSM-10 Timelines (15:30) Achieving fault-tolerant quantum computation requires bridging the gap between raw physical qubits and error-corrected logical qubits, which currently operate at an approximate 1,000-to-1 physical-to-logical ratio. While theoretical designs like Majorana and Ocelot remain in development, hardware demonstrations like Google's Willow chip represent milestones approaching the single logical qubit threshold. Algorithms like Harrow-Hassidim-Lloyd (HHL) illustrate the potential for exponential speedups in solving complex linear equations once fault tolerance is realized. Under current optimistic forecasts, asymmetric public-key cryptography could fall to quantum cryptanalysis by around 2035. This horizon underpins National Security Memorandum 10 (NSM-10), which mandates that federal systems and their commercial defense partners achieve post-quantum resistance by 2035, supported by internal federal implementation goals targeting 2030. Because AI can accelerate quantum hardware design and error correction, the timeline toward a functioning CRQC may arrive faster than historical models predict. Jones stresses that organizations cannot afford to delay migration to post-quantum standards until fault-tolerant hardware arrives, particularly given ongoing data harvesting campaigns.

精炼阅读摘要

美国国土安全部（DHS）战略技术副主管加菲尔德·琼斯（Garfield Jones）指出，人工智能与量子计算正在形成相互加速的双向融合态势；这种融合虽然有望突破大语言模型（LLM）的算力瓶颈与逻辑推理限制，但具备密码分析能力的量子计算机（CRQC）也将彻底瓦解现有的非对称加密体系，迫使防御机构必须立即防范数据窃取与对抗性AI安全风险。 AI与量子计算的融合与技术门槛 (0:04) 琼斯强调，将人工智能与量子计算结合并非即插即用，现有大语言模型完全基于经典二进制计算架构，无法直接在量子硬件上有效运行。要实现真正的量子加速AI，技术界必须跨越三项核心门槛：开发专为量子力学特性定制的专用量子LLM算法、构建量子处理器能够原生读取与处理的新型数据结构，以及推动量子硬件成熟至可稳定运行大规模模型负载的水平。此外，混合计算模式（如量子近似优化算法QAOA）能够动态分流部分计算任务，优化经典计算机处理效率较低的特定工作负载。 在应用潜力方面，这种技术整合有望为现代AI带来实质性突破。琼斯指出，当前LLM在复杂逻辑与抽象推理上存在明显的底层瓶颈，而量子计算提供的算力支持能够强化逻辑推理架构。同时，量子计算能够指数级缩减基础大模型的训练周期，并在高维数据集中实现高效、高保真的复杂合成数据生成，有效填补现实世界训练样本匮乏的空白。 CRQC对密码体系与数据的威胁 (3:35) 从网络安全防御角度来看，网络安全和基础设施安全局（CISA）等机构高度关注通用门式量子计算机所引发的CRQC威胁。一旦CRQC研制成功，彼得·肖尔在1994年提出的肖尔算法（Shor's Algorithm）将在数秒内完成大质数分解与离散对数计算，彻底摧毁现代公钥基础设施（PKI）所依赖的RSA、椭圆曲线加密（ECC）及数字签名体系。相比之下，格罗弗算法（Grover's Algorithm）仅对对称加密产生二次方加速削弱，可通过增大密钥长度进行防御。 琼斯特别警示了敌对势力目前正广泛采用的“先截获，后解密”（HNDL）策略。攻击者目前正在持续拦截并仓储机构和国家层面的加密通信与数据资产，意图在未来CRQC成熟时集中解密。这种滞后解密不仅会完全暴露历史机密，还会使机构的底层数据结构与AI训练流水线暴露于长期的系统性逆向分析和篡改风险之中。 量子增强AI带来的对抗性攻击风险 (13:42) 量子计算与生成式AI的结合在对抗环境中形成了显著的非对称优势。借助能够求解复杂优化问题的量子算法（如求解线性方程组实现指数加速的HHL算法），恶意行为者可以系统化检测并挖掘经典安全模型、算法架构与机器学习系统中的隐藏漏洞。同时，攻击者还能联合运用量子计算与AI分析多维非关联数据，绕过常规隐私保护协议，对匿名数据集实施精准去匿名化重构。 不仅如此，量子能力在AI训练管道中带来了严重的中间人攻击与数据投毒隐患。由于传输中数据和密钥交换协议普遍采用非对称加密，一旦这些通道在量子攻击下失效，恶意攻击者便可在数据进入模型前进行隐蔽篡改。琼斯警告，这种双向加速机制不仅大幅降低了高杀伤力、定制化恶意内容与网络武器的生成门槛，也让CRQC本身的实现时间表被进一步压缩。 容错硬件里程碑与后量子迁移合规 (15:30) 在硬件发展路径上，评估量子成熟度必须严格区分物理量子比特与容错逻辑量子比特。受限于极端超低温运行环境与量子相干性维持难度，构建单个纠错逻辑量子比特大约需要1000个物理量子比特的支撑。尽管马约拉纳（Majorana）和Ocelot等路径仍偏向理论构想，但谷歌Willow等物理芯片已取得显著工程进展，正逐步迈近单逻辑量子比特的临界门槛。 针对攻破公钥加密的时间预测模型，在乐观预测情景下，CRQC可能在2035年左右具备破密能力。为应对此危机，美国国家安全第10号备忘录（NSM-10）强制要求联邦机构及其私营国防商业伙伴在2035年前实现对后量子密码（PQC）的全面抗性迁移，内部执行路线图更将关键准备节点提前至2030年。美国国家标准与技术研究院（NIST）亦已正式推出首批PQC标准（FIPS 203、FIPS 204和FIPS 205），要求全行业在技术奇点出现前迅速推进防御升级。 琼斯总结指出，AI对量子工程设计与纠错机制的反哺可能会大幅压缩实现CRQC的预期时间线，使机构面临技术突袭的风险。当前的核心挑战在于，量子硬件的物理实现与逻辑纠错仍需克服严苛的工程瓶颈，而防御侧的后量子密码迁移与数据合规必须在2030至2035年的窗口期内全面铺开，以应对不可逆的存量数据破译危机。

Detailed notes — zh

## Background 该视频由人工智能会议平台 Ai4 于 2025 年 8 月 28 日发布在 YouTube 上。内容为美国国土安全部（DHS）战略技术副主管 Garfield Jones 带来的主题演讲，其摘要指出当前人工智能与量子计算正在相互促进发展，但目前关于大语言模型（LLM）与量子计算融合的大多数认知仍基于不完全的研究与理论假设。 ## TL;DR 演讲系统梳理了量子计算与人工智能交汇发展的核心路径，指出两者深度融合依赖于硬件算力跃升、专属量子算法设计以及新型数据表示三大前提。Jones 警示称，具备密码分析能力的量子计算机（CRQC）将凭借 Shor 算法瓦解现有公钥密码体系，催生“先收集后解密”（HNDL）及训练数据操纵等严峻对抗风险。与此同时，量子算力与 LLM 的共生演进有望破解模型的抽象推理瓶颈并成倍缩短训练周期，促使美国政府加速推进 NSM-10 要求、力争在 2035 年前全面完成向后量子密码（PQC）体系的迁移。 ## Atmosphere & style 整场演讲风格兼具政府安全高管的权威严谨与幽默平实的叙事口吻。演讲者在开场以轻松自嘲打破冰冷的技术距离感，随后迅速切入直击要害的网络威胁与底层数学原理，语调冷静而充满警示意味。现场没有华丽的视觉包装，更多是通过扎实的技术洞察与务实的时间表推演，营造出一种在技术突袭前夕保持清醒、紧迫却又胸有成竹的专业氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) 量子计算与大语言模型（LLM）的深度融合依赖三大发展前提：量子硬件算力达标、开发专属量子LLM算法，以及构建适配量子态的全新数据表示方式。 - [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) 量子计算研发分为专用（如聚焦优化与量子退火）与通用两条路径，其中通用量子计算催生的 CRQC（具备密码分析能力的量子计算机）是网络安全的最大威胁。 - [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) CRQC 能够利用 Shor 算法在多项式时间内破解 RSA 和 ECC 等主流公钥加密体系，彻底颠覆现有的数字签名和通信加密基础设施。 - [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) 为抵御量子解密威胁，美国国家标准与技术研究院（NIST）已正式发布三项后量子密码学（PQC）标准：FIPS 203、FIPS 204 和 FIPS 205。 - [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) 量子算力能够为大模型训练带来指数级加速，大幅缩短训练耗时，并能高维模拟复杂数据分布以生成高质量合成数据。 - [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) 在对抗加密机制时，Shor 算法对非对称加密具有指数级加速破解能力，而 Grover 算法对对称加密仅提供二次方加速，因此 AES-256 等对称加密依然相对安全。 - [11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) 量子计算强大的数学求解与优化算力，有望补齐当前经典 LLM 在抽象推理（Abstract Reasoning）和复杂逻辑推演上的性能瓶颈。 - [12:37](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=757s) “先收集后解密”（HNDL）攻击正在发生，黑客囤积海量加密数据等待 CRQC 问世后解密，进而可能实施精准的训练数据操纵和模型投毒。 - [12:37](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=757s) 结合量子算力的大模型（QC-LLM）不仅能快速挖掘算法漏洞，还可能被用于高效去匿名化私有数据并生成高破坏性的对抗攻击内容。 - [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) 根据美国 NSM-10 国家安全备忘录，联邦系统及相关供应链必须在 2035 年前全面完成抗量子密码体系（PQC）的部署与迁移。 - [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) 实现具备纠错能力的 1 个逻辑量子比特通常需要约 1000 个物理量子比特，硬件纠错与低温物理环境仍是量子计算走向实用的核心瓶颈。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) 量子计算与人工智能的交汇与发展前提 来自美国国土安全部（DHS）战略技术副主管 Garfield Jones 开场感谢了 Ai4 大会的邀请，并对现场爆满的观众人数表示惊讶。他幽默地提到：“在五年前，量子计算从来都不是什么热门或酷炫的话题，但如今情况显然已经彻底改变。”当前，各大科技巨头如 AWS、微软、IBM 都在大力推进量子领域的研发并发布新成果。Jones 特别申明，作为政府官员，他并不为任何特定供应商背书，但业界目前的研究进展确实非常可观。 针对当下的 AI 浪潮，Jones 指出量子计算与人工智能的深度融合需要跨越三个核心发展阶段： 1. **量子计算硬件能力达到能够运行大语言模型（LLM）的期望水平**：当前的大语言模型是基于经典计算体系架构设计的，无法直接在现有的量子计算机上良好运行。 2. **开发专门适配量子计算的全新 LLM 算法**：研究人员必须从底层为量子框架量身定制算法体系。 3. **构建适用于量子计算的数据表示方式**：数据是整个人工智能的基石，数据在量子环境下的输入与处理结构必须全面革新。 虽然这一融合距离全面落地仍有一段路要走，但技术演进的速度极快。Jones 回忆道，他第一次参加并演讲的量子计算会议全场只有 10 个人，而如今量子科技的发展势头已不可同日而语。 ```mermaid flowchart TD accTitle: 量子计算与AI融合的三大支柱 subgraph Prerequisites["量子与AI融合的关键发展条件"] A["量子硬件算力升级<br/>(达到运行LLM的门槛)"] B["专属量子LLM算法<br/>(替代传统经典模型算法)"] C["新型量子数据表示<br/>(适配量子态的数据结构)"] end Prerequisites --> D["量子增强型人工智能<br/>(Quantum AI)"] ``` --- ### [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) 量子计算的两大研发路径与 CRQC 的安全威胁 在梳理量子计算的技术背景时，Jones 坦言自己不擅长制作美观的图表，随后深入剖析了推动量子计算的两条核心研发路径： - **专用量子计算（Special-Purpose Quantum Computing）**：高度聚焦于特定小众领域，特别是以优化问题为核心，其中量子退火（Quantum Annealing）是目前业界非常热门的一条路线。 - **通用量子计算（Universal / General-Purpose Quantum Computing）**：能够执行通用量子逻辑门操作的计算架构。 作为来自网络安全与基础设施安全局（CISA）的技术专家，Jones 指出政府和安全部门最关注的是通用量子计算所催生的**具备密码分析能力的量子计算机（Cryptanalytically Relevant Quantum Computer, CRQC）**。CRQC 一旦问世，将彻底瓦解当今现有的公钥加密体系，包括所有私人通信、电子邮件、数字签名、RSA 算法以及椭圆曲线密码学（ECC）。 幸运的是，美国国家标准与技术研究院（NIST）具备前瞻性，已经主导制定并发布了三项后量子密码学（PQC）标准，即 **FIPS 203、FIPS 204 和 FIPS 205**。 在硬件物理实现层面，构建 CRQC 面临极高的工程挑战，需要研发特殊材料，并将系统维持在“比外太空还要冷”的超低温环境下以确保量子相干性。在数学原理上，Peter Shor 于 1994 年提出的 **Shor 算法** 证明，只要运行在足够强大的量子计算机（即 CRQC）上，就能在多项式时间内完成大整数质因数分解和离散对数计算，从而破解当前的公钥体系。此外，在计算模式上，业界也在探索混合架构（例如量子近似优化算法 QAOA），在同一计算流中根据任务特性让量子与经典处理器各司其职。 - **通用威胁**：CRQC 能够对经典公钥密码学实现指数级加速破解。 - **后量子标准**：FIPS 203（ML-KEM 密钥封装）、FIPS 204（ML-DSA 数字签名）、FIPS 205（SLH-DSA 无状态哈希签名）。 - **物理挑战**：极低工作温度、超导材料与量子比特稳定性控制。 --- ### [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) 量子算力对 AI 的赋能与新型网络对抗风险 Jones 进一步分析了 CRQC 带来的指数级速度跃升如何反哺人工智能领域。对于当前困扰 AI 开发者的超长大模型训练周期，量子计算机能够实现指数级的加速，大幅缩短训练耗时。同时，量子算力能使大模型处理更高维度、更具多样性的海量数据，全面增强模型性能；在合成数据（Synthetic Data）生成方面，量子计算能够突破样本量不足的限制，精准模拟现实复杂分布并生成高质量数据。 然而，量子与 AI 的结合也带来了严峻的“阴暗面”。攻击者同样能够利用量子计算与 LLM 对网络资产发起攻击： - **公钥/非对称加密的脆弱性**：CRQC 能够通过 Shor 算法瞬间攻破模型训练数据库的非对称加密传输与身份验证，攻击者可轻易实施中间人攻击（Man-in-the-Middle），劫持或篡改训练数据。 - **对称加密的相对安全性**：用于保护训练数据库静态存储的对称密钥对 Shor 算法免疫。要破解对称加密，量子计算机需要运行 Grover 算法，其仅能提供二次方加速（Quadratic Speedup），因此安全性相对更有保障。 - **训练数据需求量**：量子算法的加速机制可能减少对传统海量样本的依赖，但现阶段的量子计算机尚未达到实用化支撑完整 LLM 训练的成熟度。 | 密码机制类型 | 代表算法/应用 | 对应的量子破解算法 | 量子加速级别 | 安全现状评估 | |:--- |:--- |:--- |:--- |:--- | | **非对称加密 (公钥)** | RSA、椭圆曲线 (ECC)、数字签名 | Shor 算法 (1994) | 指数级加速 (Exponential) | **极度危险**，将被 CRQC 完全破解，需迁移至 FIPS PQC 标准 | | **对称加密 (私钥)** | AES 等训练数据库静态存储加密 | Grover 算法 | 二次方加速 (Quadratic) | **相对安全**，通过增加密钥长度（如 AES-256）可有效防御 | ### [11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) 量子计算赋能大语言模型：从抽象推理到共生演进 虽然当前技术尚未完全成熟，但技术突袭往往毫无征兆地发生，让人措手不及。在探讨大语言模型（LLM）的能力边界时，Garfield Jones 重点强调了“抽象推理”（abstract reasoning）。当前 LLM 在逻辑推演和复杂推理方面仍存在明显短板和偶发故障，而量子计算强大的算力支持有望补齐这一缺陷，大幅提升模型的逻辑与推理性能。 这种 AI 与量子计算之间的共生演进（symbiotic relationship）正在加速两者的技术融合。算力的指数级跃升不仅将拓展 LLM 的能力上限，也促使人们开始正视伴随而来的安全挑战。 ```mermaid flowchart LR accTitle: AI 与量子计算的共生融合 subgraph AI_Side["人工智能 (AI / LLM)"] A["大语言模型 (LLM)"] B["抽象推理与逻辑瓶颈"] end subgraph QC_Side["量子计算 (QC)"] C["指数级算力提升"] D["量子算法优化"] end subgraph Fusion["共生演进与安全威胁"] E["增强型 QC-LLM"] F["CRQC 研制加速 / 对抗风险"] end B -->|量子算力赋能| E C -->|解决优化难题| E A -->|算法与模型优化| D E --> F ``` ### [12:37](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=757s) 对抗性风险：“先收集后解密”与训练数据操纵 在带来能力突破的同时，量子技术与 AI 的结合也带来了严峻的安全隐患。最具代表性的是“先收集、后解密”（Harvest Now, Decrypt Later, HNDL）攻击：攻击者当前正在海量窃取并囤积加密数据，等待密码学相关量子计算机（Cryptanalytically Relevant Quantum Computer, CRQC）问世后集中解密。一旦加密数据被破解，攻击者便能摸清底层数据结构，进而实施精准的训练数据操纵（training data manipulation）。 此外，量子计算求解复杂优化问题的能力使其能够快速挖掘算法和模型中的漏洞。结合了量子计算能力的 LLM（QC-LLM）相比传统 LLM 具有压倒性优势，攻击者可以利用其实现私有数据的去匿名化，并以远超以往的速度和破坏力生成恶意对抗内容。 - **先收集后解密（HNDL）**：攻击者持续截获并存储敏感数据，静待 CRQC 突破以实施解密与操纵。 - **算法漏洞挖掘**：利用量子优化能力精准定位现有算法和安全架构中的脆弱点。 - **隐私与内容生成风险**：利用 QC-LLM 快速实施去匿名化攻击，高效炮制高危害性的恶意内容。 ### [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) 密码学过渡时间表、量子比特门槛与未来展望 围绕后量子密码学（PQC）的发展路径，演讲展示了涵盖悲观、历史与乐观趋势的预测模型。根据第 10 号国家安全备忘录（NSM-10）的要求，美国政府及相关供应链企业必须在 2035 年之前全面实现具备抗量子密码能力的系统部署（PQC 迁移推进目标设定在 2030 年左右）。在乐观预测下，破解 RSA 和椭圆曲线密码学（ECC）的关键节点指向 2035 年；若计入 AI 对量子研发的反哺加速效应，这一时间拐点还可能大幅提前。 在硬件进展方面，业界常讨论的量子比特多为“物理量子比特”（Physical Qubit），而构建一个“逻辑量子比特”（Logical Qubit）通常需要约 1000 个物理量子比特配合纠错机制。像 Majorana 和 Ocelot 等方案多偏向理论探索，而 Google 的 Willow 芯片则展现了更多实证成果，更接近 1 个逻辑量子比特的实用门槛。结合 Shor 算法（破解公钥体系）、Grover 算法（削弱对称加密 AES）以及 HHL 算法（Harrow-Hassidim-Lloyd，用于线性方程组的量子加速与优化），AI 与量子计算的交织正在重塑网络安全的未来格局。 - **NSM-10 政策基准**：要求联邦系统在 2035 年前全面具备抗量子攻击能力，私营部门与政府对接必须同步采用 PQC 算法。 - **物理量子比特 vs 逻辑量子比特**：形成 1 个具备纠错能力的逻辑量子比特大约需要 1000 个物理量子比特。 - **关键量子算法**： - **Shor 算法**：对 RSA 及 ECC 等公钥加密体系构成根本威胁。 - **Grover 算法**：加速对称加密（如 AES）的暴力破解。 - **HHL 算法（Harrow-Hassidim-Lloyd）**：实现线性方程组及相关复杂运算的高效量子优化。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) 在五年前，量子计算从来都不是什么热门或酷炫的话题，但如今情况显然已经彻底改变。 — Garfield Jones > [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) 量子计算硬件能力达到能够运行大语言模型（LLM）的期望水平 — Garfield Jones > [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) 开发专门适配量子计算的全新 LLM 算法 — Garfield Jones > [00:04](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=4s) 构建适用于量子计算的数据表示方式 — Garfield Jones > [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) CRQC 一旦问世，将彻底瓦解当今现有的公钥加密体系，包括所有私人通信、电子邮件、数字签名、RSA 算法以及椭圆曲线密码学（ECC）。 — Garfield Jones > [03:35](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=215s) 比外太空还要冷 — Garfield Jones > [08:03](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=483s) 用于保护训练数据库静态存储的对称密钥对 Shor 算法免疫。 — Garfield Jones > [11:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=690s) 当前 LLM 在逻辑推演和复杂推理方面仍存在明显短板和偶发故障，而量子计算强大的算力支持有望补齐这一缺陷 — Garfield Jones > [12:37](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=757s) 先收集、后解密 — Garfield Jones > [15:30](https://www.youtube.com/watch?v=aVT9PhD88Pw&t=930s) 形成 1 个具备纠错能力的逻辑量子比特大约需要 1000 个物理量子比特。 — Garfield Jones ## Mentioned Garfield Jones, Peter Shor, 美国国土安全部 (DHS), 网络安全与基础设施安全局 (CISA), 美国国家标准与技术研究院 (NIST), AWS, 微软 (Microsoft), IBM, Google, FIPS 203 (ML-KEM), FIPS 204 (ML-DSA), FIPS 205 (SLH-DSA), 第 10 号国家安全备忘录 (NSM-10), CRQC (具备密码分析能力的量子计算机), Shor 算法, Grover 算法, HHL 算法, QAOA (量子近似优化算法), Willow 芯片, Majorana, Ocelot, RSA, 椭圆曲线密码学 (ECC), AES, 先收集后解密 (HNDL) ## Video description Garfield Jones, Associate Chief of Strategic Technology, U.S. Department of Homeland Security Abstracts There has been increasingly more research conducted on ways AI can help advance quantum computing (QC) and vice-versa. Most QC information regarding large language models (LLMs) is based on incomplete research or hypotheses of researchers. [Source](https://www.youtube.com/watch?v=aVT9PhD88Pw)

Source quotes

[0:04] Garfield Jones: 在五年前，量子计算从来都不是什么热门或酷炫的话题，但如今情况显然已经彻底改变。 [0:04] Garfield Jones: 量子计算硬件能力达到能够运行大语言模型（LLM）的期望水平 [0:04] Garfield Jones: 开发专门适配量子计算的全新 LLM 算法 [0:04] Garfield Jones: 构建适用于量子计算的数据表示方式 [3:35] Garfield Jones: CRQC 一旦问世，将彻底瓦解当今现有的公钥加密体系，包括所有私人通信、电子邮件、数字签名、RSA 算法以及椭圆曲线密码学（ECC）。 [3:35] Garfield Jones: 比外太空还要冷 [8:03] Garfield Jones: 用于保护训练数据库静态存储的对称密钥对 Shor 算法免疫。 [11:30] Garfield Jones: 当前 LLM 在逻辑推演和复杂推理方面仍存在明显短板和偶发故障，而量子计算强大的算力支持有望补齐这一缺陷 [12:37] Garfield Jones: 先收集、后解密 [15:30] Garfield Jones: 形成 1 个具备纠错能力的逻辑量子比特大约需要 1000 个物理量子比特。

Source quotes

[0:04] Garfield Jones: Current large language models (LLMs) are architected exclusively around classical computing frameworks and will not run effectively on existing quantum systems. [3:35] Garfield Jones: A sufficiently powerful CRQC has the capability to dismantle the entire foundation of modern public-key infrastructure (PKI), rendering standard asymmetric encryption methods—including RSA, elliptic-curve cryptography (ECC), and digital signatures across emails and private messaging—entirely obsolete. [8:03] Garfield Jones: Threat actors equipped with quantum capabilities could exploit asymmetric vulnerabilities to execute undetectable man-in-the-middle attacks, intercepting, poisoning, or manipulating AI training pipelines and access pathways before the data ever reaches the model. [11:30] Garfield Jones: Hostile actors continuously siphon encrypted institutional and national data, warehousing it until a Cryptanalytically Relevant Quantum Computer (CRQC) becomes operational. [13:42] Garfield Jones: By leveraging quantum algorithms capable of solving complex optimization problems, adversaries can systematically expose hidden vulnerabilities within classical algorithms, defense frameworks, and machine learning models. [13:42] Garfield Jones: This bidirectional dynamic—where AI accelerates quantum computing breakthroughs and quantum computing turbocharges AI capabilities—drastically compresses the timeline toward achieving a fully functional CRQC. [15:30] Garfield Jones: NSM-10 mandates that government systems transition to Post-Quantum Cryptography (PQC) resistance, an obligation that directly extends to private-sector vendors doing business with the federal government, with internal implementation roadmaps pushing aggressively toward a 2030 target. [15:30] Garfield Jones: Hardware evaluations must account for logical rather than raw physical qubits, operating at a ~1,000:1 physical-to-logical ratio.

Original transcript

[0:04] First of all, AI4, thanks for having me [0:06] here. Um, thank you for coming in to to [0:10] hear the talk here. I'm actually really [0:12] surprised at how many people are are in [0:14] because um, I always say quantum is [0:17] never it was never the popular or the [0:19] cool topic, you know, 5 years ago, but [0:22] of course things have changed. Um [0:26] so as we go into this um I'm going to [0:29] talk a little bit about you know a [0:31] little bit about the background uh about [0:33] the research that's starting to you know [0:37] really [0:38] increase in in the quantum space as [0:42] you're probably seen you know AWS [0:44] Microsoft [0:46] um you know IBM of course they're all [0:49] pushing uh a lot of the they they all [0:52] have these these releases. I'm not [0:54] endorsing it. I am from uh the [0:56] government, so I don't endorse any any [0:59] vendors, but they there's a lot of good [1:01] research in there. Um you know, as you [1:05] as you look at AI, and this is a AI, you [1:08] know, conference. um there's some [1:11] developments that have to kind of really [1:14] happen for AI and quantum to really [1:16] merge and to for quantum to really have [1:19] a or quantum computing to really have an [1:21] effect on artificial intelligence. So as [1:24] we we look here there really you know [1:26] three primary developments and as you [1:28] look at the slide and I'm sorry if you [1:30] can't see from the back so I'll I'll [1:31] just kind of read them out. Um I was [1:34] back there and I I saw the slides and [1:35] you can barely see them. Um there are [1:38] three developments that you you really [1:40] have to look at when quantum computing [1:42] can execute an LLM in a desired level or [1:45] capability and quantum computing of [1:47] course is a little different than the [1:49] the classical computing. So an LL an LLM [1:52] or a large language model will not [1:54] necessarily run really well on quantum [1:57] computers that we have now. So there's [1:59] there's going to be some time uh new uh [2:02] large language models or LLM algorithms [2:05] tailored for quantum computing have to [2:08] be developed. So there's there's some [2:09] time for that and and there's some [2:11] research that's already going into that. [2:13] So that's something that's that's really [2:15] important that we understand the the the [2:18] models that we have now are are are [2:20] geared towards that classical computing [2:23] and they're focused on that. So as we go [2:25] to towards the quantum computing and [2:27] I'll give you a little bit of background [2:28] on quantum computing in in a couple [2:30] slides but um that's a it's a little [2:33] different uh framework that you have to [2:35] work with in in terms of uh running the [2:37] the algorithms that are developed for uh [2:40] the classical um the the classical [2:43] computing side. Um of course my phone [2:46] would go dead. Um [2:48] uh so you also have new data [2:51] representation that needs to be suitable [2:53] for quantum usage. So that's another [2:56] piece that that really has to be be [2:58] pushed on there. So there's there's a [3:00] lot of good stuff that'll come out when [3:03] quantum computing meets AI but we have [3:05] to design that those uh the what the [3:09] foundation of all AI which is the is the [3:11] data and that has to be um developed as [3:14] well. So as as we can see that it it's a [3:18] little ways out but um you know I never [3:21] say anything is too far out because uh [3:24] the first quantum computing conference [3:25] that I went to and and spoke at had 10 [3:28] people in it and so it's it's moved [3:31] quite quite quickly. [3:35] Uh so as we we look here we got we've [3:38] got um a little bit of background on [3:41] quantum computing. So you have the the [3:43] two types of quantum computing and I'm [3:45] sorry for the tech heavy stuff. I am not [3:47] a graphical artist. Uh matter of fact I [3:50] suck at graphics. Um but so you have two [3:54] R&D approaches that really can kind of [3:58] drive uh quantum computing and one is [4:01] the special purpose computing or special [4:04] purpose quantum computer which is very [4:06] it's very niche. It's very um you know [4:08] it's very focused on optimization which [4:10] quantum computing is is uh is heavily [4:13] based and then of course the quantum [4:16] analing which I think is one of the more [4:18] popular uh avenues that that folks take [4:21] in in quantum computing. The other piece [4:23] is the universal or general purpose [4:25] quantum computing and that is where [4:27] we're really concerned coming from the [4:30] cyber security infrastructure security [4:31] agency. That's that's the that's where [4:34] we're looking at as far as where that [4:37] cryptographically quantum computer or [4:39] the cryptoanalytically quantum computer [4:42] um will will really come come to play. [4:45] So you'll hear the term CRQC and a lot [4:49] um and that's what we're really [4:51] concerned about and that CRQC is is uh [4:56] you know geared towards that that um [4:58] general purpose quantum computing. So [5:00] that's where we're really um focused on. [5:03] And that CRQC can break all the public [5:07] key encryption that we have today. All [5:09] your private messages, all your your [5:12] email, your signatures, everything. So [5:15] RSA, um curve, all the digital [5:18] signatures. Thank goodness for NIST that [5:20] they saw this uh you know they they had [5:23] the foresight to start working on the [5:25] algorithms. So they developed the three [5:27] algorithms that we have today that are [5:30] um PQC or postquantum cryptography um or [5:34] quantum resistant. So those are if [5:36] you're not familiar with them, they're [5:37] they're FIPS 203, 204, and 205. I always [5:41] forget the name, so just just take my [5:44] word on it. It's FIPS um certified and [5:48] it's 203, 204, and 205. [5:51] Um so there's a lot of R&D on special [5:55] purpose of course uh we we're seeing [5:57] that but you know breakthroughs happen [6:01] and so once you never know what the [6:04] special purpose the special purpose [6:06] could um push the uh CRQC level and that [6:11] CRQC there's a lot of stuff that that [6:13] has to happen with that. There's got to [6:15] be the right materials. um it's it's got [6:18] to be colder than space, you know, to to [6:20] keep that those mathematical um [6:23] computations going. The factoring part [6:26] of it, I mean, with all the PKI that we [6:29] have now and RSA and and elliptic curve, [6:33] those [6:35] the factoring um that we use for for [6:38] encryption can be broken by an algorithm [6:42] we call Shores algorithm, S. [6:45] And it's shor Peter Shore developed it [6:47] in 1994 and it's basically uh if it was [6:51] able to run on a sufficiently powerful [6:53] quantum computer which is the CRQC [6:56] you'll it'll be able to break all the [6:58] encryption. So that's where we're really [6:59] concerned about and of course that's on [7:01] the general purpose side. We see a lot [7:03] of work on the hybrid side as well. Um [7:07] and this is where I'm talking about [7:10] computing and not necessarily [7:11] encryption. Um so encryption there is a [7:14] hybrid part of encryption where you have [7:17] the classical and you have the um [7:19] quantum uh the quantum resistant [7:21] encryption that's being used. So there's [7:24] there's work in that. But when you're [7:26] talking about hybrid, you're talking [7:28] about the ability for and and you'll see [7:31] it there where we're talking about the [7:32] AOA to quantum AOA. Um those are [7:36] algorithms that they they do a certain [7:38] amount of quantum calculations and then [7:41] they flip over to classical to to comp [7:44] complete some of the the classical um [7:47] computations cuz there's still there's [7:49] still some problems on the classical [7:50] side that there's still some problems [7:53] that are designed for classical work and [7:55] they are they're more optimum on a [7:57] classical computer than they are on a on [7:59] a quantum computer. [8:03] So why is this the uh why is this the [8:06] problem child the the CRQC the crypto [8:09] analytically um relevant quantum [8:12] computer? So as again it breaks all the [8:15] the the uh encryption that we have [8:18] today. uh as you can see on the graph [8:20] there, we've got the how long it takes [8:22] the the best classical algorithm to go [8:25] and we get an exponential speed up from [8:28] from the CRQC and it's able to break uh [8:32] the encryption that we have very very [8:34] quickly. Um you know as we as we look at [8:37] the merge between AI and and quantum we [8:41] start to see what things that quantum [8:44] computer can really bring to artificial [8:46] intelligence. And of course the training [8:48] time, right? We we decrease the the [8:51] training time for for LLMs. And that's [8:54] that's where a lot of us are probably [8:56] working right now is like, oh man, it [8:58] takes forever to train these these large [9:00] language models. But with a quantum [9:02] computer, it could be exponentially [9:05] faster. And then there's also increased [9:07] capabilities for for that. when you're [9:10] able to put them uh a variety of data [9:13] more data you can also get more [9:15] capabilities out those LLMs so of course [9:18] quantum will will bring that space um [9:22] the other piece of it would be if you're [9:25] in the in in the um synthetic data right [9:30] that is that's really important [9:32] synthetic data sometimes is a little [9:34] hard to generate right and hard to mimic [9:37] the the real data or to if you only have [9:39] a little bit or or a small amount of [9:41] data, then you you you don't necessarily [9:44] have that um that variety or that [9:47] capability. So, having that that ability [9:50] to to um push the synthetic data is is [9:54] another piece of it. So, that's that's [9:56] where um I see that these things can [9:59] really help out. Of course, we'll talk [10:01] about the ugly where we're talking about [10:03] the adversary can use um quantum [10:06] computing and LLMs as well. And we've [10:09] seen that many times. So, it as we [10:12] talked about the CRQC can defeat the the [10:15] encryption and the training data. Um we [10:18] we look at this here and we talk about [10:21] decreased training time. um you know [10:23] access to the training base databases is [10:26] is probably you know symmetric key and [10:29] it's not necessarily vulnerable to CRQC [10:31] right now because we have um CRQC you [10:34] need to run a different algorithm not [10:36] shores but you would need to run Grovers [10:39] to really break the symmetric key and so [10:41] that would be something that that we're [10:43] looking at so for for here um you know [10:48] when when you're populating the the [10:50] training database that's something that [10:51] we're really worried about as far as the [10:54] asymmetric. So, it's CRQC's break break [10:57] the asymmetric and then of course the [10:59] man in the middle, you know, kind of [11:00] jumping inside there to to really be [11:03] able to to mess with with both cases as [11:06] far as training and access to the [11:08] training data. So, um as as we say, you [11:13] know, we're we're also talking about the [11:14] amount of training data needed. So, do [11:17] you need a whole bunch of training data? [11:19] Not really. Because um the the quantum [11:22] computer can can go that much faster. Do [11:25] we have the quantum computers capable of [11:27] it today? Not really. Uh there's there's [11:30] we're not there yet. But again, we we've [11:33] seen things just that technological [11:36] surprise pop out of nowhere and we we're [11:39] we're caught behind the eight-ball. So [11:41] there's there's also that is is is [11:44] really important. And as I talk about [11:45] the, you know, the the capabilities of [11:48] the LLM, um, you know, going to the the [11:52] abstract reasoning, that's where I I [11:55] really want to jump to is that abstract [11:57] reasoning. We we have some issues or [12:00] some hiccups with the LLMs right now, [12:02] right? Everyone can can kind of um nod [12:05] to that. But I think with a with the [12:08] with the help of a quantum computer, [12:10] we're going to have a lot less to worry [12:13] about in that sense and a lot more on [12:15] the on the reasoning side and the [12:17] logical side. And that's what what [12:19] quantum computing can bring to LLMs and [12:22] and that that symbiotic relationship [12:25] where you're looking at AI and and uh [12:28] quantum computing really starting to [12:29] merge because if you're giving it more [12:31] computing power obviously you're you're [12:34] going to get more and more abilities. So [12:37] um training data manipulation. So we're [12:40] we're going into the bad. A lot of [12:42] people like bad news, right? But, you [12:44] know, there's some good news, but you [12:45] know, the bad news is always the most [12:47] exciting part. So, as as we go towards [12:51] that that quantum computing getting [12:53] integrated with LLMs, we're we're [12:55] starting to see that, you know, training [12:57] data has more manipulation. You have [12:59] more ability to manipulate that training [13:01] data. And why this is important is [13:03] because today we have the harvest now [13:05] decrypt later, right? Adversaries are [13:09] constantly taking our data. They're [13:11] constantly taking our data. They're [13:13] holding it until they get a a CRQC. When [13:17] they get that CRQC, they're able to see [13:20] what our data is our our data looks [13:22] like. Obviously, if you know what our [13:24] data looks like, then you know how to [13:27] address any issues, how to manipulate [13:30] it. So there there are issues that can [13:33] can come from that. That's why we have [13:35] to also tackle that that harvest now and [13:38] decrypt later problem today. Okay. Um [13:42] there's also exploiting those [13:44] algorithmic uh vulnerabilities, right? [13:47] We have those uh the QC can really solve [13:50] complex optimization problems which you [13:53] know now you're you're looking at the [13:56] ability to generate a lot of weaknesses [14:00] towards our whatever we have right [14:03] whatever we have in place and and so and [14:06] whatever we have in in in algorithms or [14:09] models. So we have to be able to to push [14:11] that. And then of course um the quantum [14:16] computer they they definitely a QC LLM [14:20] definitely has advantages over that [14:22] classical LLM. And then of course you [14:24] know looking at the adversary using that [14:26] QC and LLM to deanonymize private data. [14:31] You know like I said they're breaking [14:33] that and then the ability to actually [14:36] generate [14:37] malicious content. so much faster, so [14:40] much more potent uh can an adversary do [14:44] can an adversary go and develop. But [14:47] that's those are some of the things that [14:49] we really have to be careful and and [14:51] watch for uh when we're looking at [14:54] combining quantum computing and and AI. [14:58] quantum computing and AI, you know, it's [15:00] a it's a great relationship, but what I [15:02] what I feel and and what I've talked to [15:04] some of the experts is once you start [15:06] merging that, you start getting closer [15:08] and closer to that the ability to [15:11] actually develop a a CRQC because there [15:14] going to be some benefits of AI that [15:15] help quantum computing and there's a [15:17] heck of a lot of benefits that that um [15:20] quantum computing can help on AI. So as [15:22] as we start to move towards that, just [15:25] think about the risks and and other [15:27] issues that we have to deal with on [15:29] that. And this is my last slide, but I I [15:33] know it's a okay, it's my my my attempt [15:35] at graphics. This is why I don't do it. [15:37] Um and so uh you know, if you look at [15:40] this here, um we have a pessimistic, we [15:44] have a a historical, and we have a um [15:47] optimistic line. And the optimistic line [15:49] is the steepest line you can see there. [15:51] And then where we break RSA and and all [15:54] the elliptic curve and everything else [15:56] is is is at 2035, right? Uh 2035 is when [16:00] NSM10 [16:02] uh says that we need to be PQC [16:04] resistant. We need to have things in [16:06] place. Uh the government needs to have [16:09] everything in place. Which means in the [16:10] private sector, you're going to have to [16:12] have PQC resistant algorithms to deal [16:14] with the government. I'll be long [16:16] retired then, but you'll be dealing with [16:18] somebody else. Um, and then if you look [16:21] here, um, you know, we we I've put a [16:26] couple of things that you've probably [16:28] seen in the news, Majeran and Ocelot, [16:30] and then Willow is actually in there um [16:33] pretty close to the the logical cubit. [16:36] The reason I say Majerano and Ocelot, [16:38] and I'm not endorsing anyone, and the [16:41] reason it's so low on on there, it's [16:42] because they're they're theoretical [16:44] terms, right? They're theoretical [16:46] approaches. Willow was more it it's a [16:49] little it demonstrated a little bit more [16:51] and that's why we put it closer to the [16:55] uh the the one logical cubit uh [16:59] threshold and that's really important [17:01] because you you hear a lot of times [17:02] about cubits and that's physical cubits. [17:05] Logical cubits are about a thousand [17:07] physical cubits to to one logical cubit. [17:10] So always make sure you you ask what [17:13] what's going on as you see the the the [17:15] pessimistic line or I'm sorry the [17:18] optimistic line where I talked about [17:20] Shor's algorithm right as we look at at [17:23] and this was not taking AI into account. [17:26] So, as we look at more influence that AI [17:29] has on on quantum, you're going to see [17:32] that line kind of curve a little bit [17:35] towards me, right? Because um things are [17:39] going to speed up significantly, right? [17:41] So, there's that is um our best [17:45] guesstimate at what it will look like [17:47] for now. As you see, the optimistic um [17:50] view is is is still 2035 and that's [17:53] where we have to be prepared for. We're [17:55] we're on a schedule. We're pushing [17:56] toward get every everyone on 2030. But [17:59] as we go towards towards that, we want [18:02] to make sure that um we don't uh get [18:05] surprised and that line just all of a [18:08] sudden flips towards me and then we're [18:10] we're in trouble. A couple of the other [18:12] things I talked about Grovers, you know, [18:14] breaking AES and HHL, which is like what [18:17] is that the Harrow Hasslam uh Lloyd [18:21] algorithm, which is basically being able [18:23] to break all um not break but optimize [18:27] uh linear type equations and and or [18:31] linear type work. And so you're you're [18:34] looking at some some really amazing [18:36] things that that the the relationship [18:39] between AI and quantum can really build [18:41] up. So um that's kind of our our um our [18:46] thoughts on this and you know u thank [18:48] you guys for your your attention. [18:50] Hopefully uh it was informative to you [18:52] and and please uh let me know if you [18:55] have any any feedback or any um input on [18:57] it and uh I'll be around. Thank you.