Sola Security Abandons LLM Defenses for Graph-Based World Models
Ai4
Language models fail to scale in cybersecurity because enterprise networks operate as interconnected graphs rather than linear text, requiring JEPA-style world models to achieve autonomous threat remediation without operational outages.
## Background
Published on YouTube by the Ai4 channel on August 19, 2026, this presentation features Guy Flechter, Co-Founder and CEO of Sola Security. According to the video description, the session examines the core limitations of Large Language Models (LLMs) in cybersecurity and discusses architectural solutions to overcome them.
## TL;DR
Guy Flechter argues that LLMs cannot scale for enterprise cybersecurity because security is fundamentally an interconnected graph problem rather than a linear language or text problem. Current agentic LLM loops plateau at an unreliable 80%–85% success rate due to severe context bottlenecks and hallucination risks in dynamic enterprise environments. To achieve autonomous defense, Flechter proposes Cybersecurity World Models leveraging JEPA architectures that model the deterministic 'physics' of digital assets, identities, and permissions to simulate and predict remediation outcomes in a single shot without breaking operational workflows.
## Atmosphere & style
The presentation carries a pragmatic, technically grounded tone that balances candid critiques of current generative AI hype with an authoritative vision for the future of cyber defense. Flechter speaks in an engaging and conversational manner, leaning into relatable industry maxims and audience interactions during the Q&A. The session feels like an intellectual deep dive for security practitioners, moving methodically from the structural flaws of today's toolsets to the pioneering engineering required to build predictive world models.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) LLMs face fundamental scaling limitations in cybersecurity because enterprise security is an interconnected graph problem rather than a linear language problem.
- [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) Defenders and LLMs traditionally think in lists and text, whereas attackers exploit interconnected graph relationships across assets, identities, and network paths.
- [02:18](https://www.youtube.com/watch?v=px24wf8UtPM&t=138s) Current AI-native security tools primarily use LLMs as natural language translators to generate deterministic queries (SQL, Python, Cypher), which fails to provide true autonomous reasoning.
- [02:18](https://www.youtube.com/watch?v=px24wf8UtPM&t=138s) Context injection and agentic loops plateau at an 80% to 85% success rate, leaving a 15% to 20% margin of error that makes them untrustworthy for autonomous production defense.
- [04:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=285s) Cybersecurity requires foundational World Models borrowed from physical AI that understand environmental dynamics and predict the outcomes of actions instead of relying on memorized text.
- [04:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=285s) Using architectures like Yann LeCun's JEPA, cyber world models learn the underlying 'physics' of enterprise graphs to evaluate state changes without hallucination.
- [09:12](https://www.youtube.com/watch?v=px24wf8UtPM&t=552s) A world model enables single-shot blast-radius quantification during misconfigurations, instantly predicting whether an attacker can reach production databases.
- [09:12](https://www.youtube.com/watch?v=px24wf8UtPM&t=552s) Predictive world models solve remediation risk by simulating access policy changes beforehand to verify they will not cause legitimate service outages.
- [13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) Data engineering is the primary training bottleneck; models are trained by generating thousands of synthetic enterprise worlds and predicting masked subgraphs and transitions.
- [13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) The digital equivalent of physical laws consists of deterministic security constraints, such as immutable permission barriers, identity relationships, and authentication controls.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) The Core Limitation: Cybersecurity as a Graph Problem vs. Language Lists
Guy Flechter introduces Sola Security’s thesis developed over the past two years, stating upfront that despite their general capabilities—joking that he hopes no one from OpenAI or Anthropic is in the room—Large Language Models (LLMs) face hard limitations and will not scale for cybersecurity. To frame the fundamental issue, Flechter references a classic security adage from the creator of BloodHound (an open-source Active Directory reconnaissance tool): *"defender thinks in lists, attacker thinks in graphs. As long as this is true, attackers will win."*
Flechter argues that LLMs reproduce the exact mindset of the defender: they treat cybersecurity as a language problem, processing information in lists and text. While code analysis can genuinely be treated as a language challenge, the broader enterprise environment is not a text problem. Whether an organization has 20 employees or tens of thousands, enterprise security is a complex graph consisting of deeply interconnected assets, identities, permissions, services, and network paths.
```mermaid
flowchart LR
accTitle: LLM Approach vs Enterprise Reality
subgraph TraditionalLLM["Current LLM Paradigm"]
direction TB
A["Natural Language Prompt"] --> B["Linear Context / Lists"]
B --> C["Brittle Queries / Translators"]
end
subgraph EnterpriseGraph["Enterprise Reality"]
direction TB
D["Identities & Permissions"] <--> E["Assets & Databases"]
E <--> F["Cloud Services & Network Paths"]
end
```
[02:18](https://www.youtube.com/watch?v=px24wf8UtPM&t=138s) The Commodity LLM, Context Bottlenecks, and the Failure of Agentic Loops
The ultimate objective—the "holy grail" of cybersecurity—is building fully autonomous defense systems capable of connecting to an environment, understanding it independently, reasoning about risk, and taking protective action. Today's AI-native platforms attempt to reach this goal by using LLMs as natural language translators that convert plain text into deterministic queries such as SQL, Python, or Cypher. However, LLMs cannot navigate enterprise environments autonomously; they require an intermediary layer to bridge the gap.
As industry leaders like Microsoft's CEO have highlighted, foundation LLMs are increasingly a commodity accessible to anyone, making the context layer the true bottleneck. Providing an LLM with sufficient context to understand how an environment operates and what matters within it remains an uphill battle. Industry attempts to solve this via "Agentic Loops"—orchestrating tools and feeding deterministic context to LLMs—consistently plateau at an 80% to 85% success rate. An autonomous defense mechanism cannot be deployed or trusted in a production enterprise with a 15% to 20% margin of failure.
- LLMs currently function as translation mechanisms converting user prompts into SQL, Python, or Cypher queries.
- Context injection struggles to convey dynamic environment topology, leaving LLMs prone to hallucinations.
- Agentic loops fail to overcome deterministic limits, capping automated reliability around 80%–85%.
[04:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=285s) Introducing Cybersecurity World Models: From Physical AI to Enterprise Physics
To move past brittle LLM prompting, Flechter presents the need for a foundational Cybersecurity World Model. Borrowed from physical AI disciplines such as autonomous vehicles and robotics, world models enable machines to predict what will happen next when taking an action. An autonomous car navigating a new route encounters a completely novel environment every trip, yet safely executes a right turn because it understands the underlying physics of motion, velocity, and steering angles rather than relying on memorized text.
```mermaid
flowchart TD
accTitle: World Model Prediction Architecture
subgraph InputState["Current Environment"]
A["Enterprise Graph State: X"]
end
subgraph WorldEngine["World Model Physics Engine"]
B["JEPA Architecture / World Intelligence"]
C["Simulate Action: S3 Config / Permission Change"]
end
subgraph FutureState["Predicted Outcome"]
D["Next State Prediction: Prevent Breach"]
E["Verify: Zero Service Outage"]
end
A --> B
B --> C
C --> D
C --> E
```
In cybersecurity, a world model replaces multi-query generation and linguistic reasoning with native graph physics. Drawing on architectures like JEPA (Joint Embedding Predictive Architecture)—developed by Yann LeCun at Meta and advanced through initiatives like AMI—a cyber world model learns the "physics" of enterprise graphs. Instead of guessing or generating queries through trial and error, the model evaluates state $X$, simulates an action (such as modifying an Amazon S3 bucket policy), and predicts the exact future state to determine whether that intervention neutralizes the threat without breaking business workflows.
- World models shift AI from a linguistic translator to a predictive physics engine that encodes enterprise graphs.
- Prediction replaces trial-and-error query execution, eliminating hallucinations in operational defense.
- Foundational architectures leverage JEPA-style predictive representations rather than autoregressive token generation.
[09:12](https://www.youtube.com/watch?v=px24wf8UtPM&t=552s) Use Cases: Single-Shot Analysis, Cloud Remediation, and Preventing Outages
A functional cybersecurity world model unlocks fully autonomous self-healing environments. Connected systems can discover enterprise infrastructure, map topological relationships, and remediate vulnerabilities in a single shot without requiring humans in the loop to double-check hallucinations.
Flechter highlights two primary scenarios to illustrate the difference between current LLM pipelines and world models:
1. **Cloud Misconfiguration & Blast-Radius Quantification:** When an open S3 bucket or missing access control is detected, security teams must immediately evaluate if an attacker can reach production databases. Current LLM workflows require stitching together extensive logs, graph queries, and deterministic scripts. A world model evaluates the scenario in a single step, predicting attack reachability across the graph and executing the optimal mitigation instantly.
2. **Remediation Impact Analysis:** When altering an access rule or configuration, security teams obsess over whether the change will block legitimate microservice traffic and cause an outage. Standard practices involve tedious log analysis and manual testing. A world model predicts the operational outcome on the graph beforehand, confirming whether legitimate traffic will remain uninterrupted prior to applying the fix.
[13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) Q&A: Training Architectures, Data Preparation, and Defining the "Physics" of Cyber
During the audience Q&A, an attendee asks how one can benchmark and validate predictive cybersecurity models when future outcomes cannot easily be compared against an obvious ground truth. Flechter explains that the primary barrier is data engineering rather than compute costs:
> "It's not like taking LLM bringing billions of data points or scraping the internet and start putting weights inside the model. You need to create a lot of types of worlds."
Training utilizes a JEPA-inspired masking technique: the system constructs thousands of distinct enterprise "worlds" as structured graphs of users, devices, services, permissions, activities, and policies. It then masks or hides sections of a graph and tasks the model with predicting the missing subgraphs and transitions. While model training itself is the most compute-expensive phase, synthesizing and structuring high-fidelity enterprise environments is the most difficult challenge.
A second audience member asks what constitutes the equivalent of Newton's laws or Maxwell's equations in the digital domain, given how dynamic networks are. Flechter notes that Sola Security published research on "AI-native assets" to address this challenge. In cybersecurity, fundamental rules act as physics: a lack of permission is an immutable barrier, and missing two-factor authentication (2FA) creates an absolute path of exposure. Layering concrete security controls, identity relationships, and network topologies provides the mathematical ground truth necessary to simulate and predict digital environments reliably.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) defender thinks in lists, attacker thinks in graphs. As long as this is true, attackers will win. — Guy Flechter
> [13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) It's not like taking LLM bringing billions of data points or scraping the internet and start putting weights inside the model. You need to create a lot of types of worlds. — Guy Flechter
## Mentioned
Guy Flechter, Sola Security, OpenAI, Anthropic, BloodHound, Microsoft, Yann LeCun, Meta, AMI, Joint Embedding Predictive Architecture (JEPA), Amazon S3, Python, SQL, Cypher, Active Directory
## Video description
Guy Flechter - Co-Founder and CEO, Sola Security
We will discuss the limitations of LLM for Cybersecurity and what possible ways to overcome them.
Sola Security放弃LLM防御，转向基于图的世界模型
语言模型在网络安全领域无法有效扩展，因为企业网络是相互连接的图而非线性文本；这需要借助JEPA风格的世界模型来实现自主威胁修复，且不会引发业务中断。
Guy Flechter指出，LLM在企业网络安全中无法实现有效扩展，因为安全从根本上是一个相互关联的图问题，而非线性的语言或文本问题。在动态的企业环境中，受严重的上下文瓶颈与幻觉风险影响，当前的智能体LLM循环成功率停滞在不可靠的80%至85%。为了实现自主防御，Flechter提议采用基于JEPA架构的网络安全世界模型，对数字资产、身份和权限的确定性“物理规律”建模，从而在不破坏业务工作流的前提下，单次即可模拟并预测修复结果。
LLM在网络安全领域面临根本性的扩展限制，因为企业安全是一个相互关联的图问题，而非线性的语言问题。 (0:04)
防御方和LLM传统上倾向于以列表和文本的形式思考，而攻击者则会利用跨资产、身份和网络路径的互联图关系。 (0:04)
目前的原生AI安全工具主要将LLM用作自然语言翻译器来生成确定性查询（SQL、Python、Cypher），但这无法提供真正的自主推理能力。 (2:18)
上下文注入和智能体循环的成功率停留在80%至85%的平台期，留下了15%至20%的容错缺口，导致其无法在生产环境中承担值得信赖的自主防御工作。 (2:18)
网络安全需要借鉴物理AI的基础世界模型，这类模型能够理解环境动态并预测行动后果，而非依赖死记硬背的文本。 (4:45)
Reading summary
Large language models will not scale to solve enterprise cybersecurity because they approach environments as linear text lists rather than relational graphs, according to Sola Security co-founder Guy Flechter. Flechter contends that autonomous defense requires replacing text-generating agents with predictive cybersecurity world models that simulate digital physics.
The Graph Reality Versus Text Lists (0:04)
Flechter frames the core limitation of LLMs using the security adage that defenders think in lists while attackers think in graphs. While code analysis functions as a language challenge, enterprise environments are complex graphs composed of interconnected assets, identities, permissions, services, and network paths. Current AI platforms largely use LLMs as linguistic intermediaries to convert prompts into deterministic queries like SQL, Python, or Cypher, failing to capture holistic environment topology.
Reliability Plateaus in Agentic Loops (2:18)
As base foundation LLMs commoditize, context ingestion becomes the operational bottleneck. Flechter observes that industry attempts to orchestrate automated tools via agentic loops consistently plateau at an 80% to 85% success rate. In a production enterprise, a 15% to 20% margin of error remains too hazardous to deploy autonomous interventions without human oversight.
Adopting World Models and Digital Physics (4:45)
Borrowing from physical AI in robotics and autonomous vehicles, Flechter advocates for cybersecurity world models based on frameworks like Yann LeCun’s Joint Embedding Predictive Architecture (JEPA). Rather than relying on autoregressive token generation, a world model learns the structural physics of enterprise graphs. Immutable security controls—such as access permissions, network boundaries, and multi-factor authentication—function as digital laws of physics, allowing a system to simulate actions and predict future states without hallucinations.
Single-Shot Remediation and Synthetic Training (9:12)
In production, predictive world models are intended to enable single-shot incident remediation. When evaluating a misconfigured Amazon S3 bucket, a world model assesses attacker reachability across the graph and verifies that permission changes will not break legitimate microservice traffic or cause operational outages. Flechter notes that training these models relies on masking techniques across thousands of structured enterprise environments, arguing that synthesizing high-fidelity digital worlds presents a greater barrier than compute costs.
Flechter concludes that reliable autonomous defense cannot be achieved through commodity LLM translation layers due to inherent reliability ceilings. The primary obstacle to realizing alternative world models lies in the rigorous data engineering required to synthesize and benchmark realistic enterprise graphs.
精炼阅读摘要
Sola Security联合创始人Guy Flechter提出，大语言模型（LLM）因将互联互通的企业安全图谱误构为线性文本处理问题，无法支撑起可靠的自动化网络安全体系；唯有借鉴实体AI与JEPA架构构建网络安全世界模型，方能实现具有确定性的自主防御。
图结构对抗与语言模型的线性局限 (0:04)
Guy Flechter援引开源Active Directory侦察工具BloodHound作者的名言强调，“防守方以列表思考，攻击者以图谱思考；只要这一现状不变，攻击者就将持续获胜”。Flechter指出，大语言模型正在复制传统防守方的这一思维弊端，将网络安全视作自然语言与列表处理任务。虽然代码分析在一定程度上属于语言范畴，但现代企业环境无论规模是20人还是数万人，本质上都是由高度交织的资产、身份凭证、访问权限、云服务以及网络路径构成的复杂图谱（Graph）。
现存LLM架构依赖线性上下文与文本提示词，无法自然映射网络拓扑结构。攻击者擅长沿着图谱中的非线性路径实施权限提升与横向渗透，而LLM只能将信息扁平化为列表进行脆弱的查询转换，这种范式层面的不对称导致大模型在面对复杂的企业攻击路径时难以进行有效推演与全局防御。
上下文瓶颈与智能体循环的可靠性上限 (2:18)
自动化网络防御的终极目标是建立能够自主接入生产环境、独立理解风险并实施阻断的系统。Flechter分析称，当前业界AI原生方案普遍仅将基础大语言模型作为翻译中介，把自然语言指令转化为SQL、Python或Cypher等确定性查询语句。然而，随着基础LLM快速商品化，真正的技术瓶颈转移到了动态环境的上下文注入层，向模型完整传递拓扑状态与业务资产权重的工程难度极高。
为了突破确定性环境与LLM之间的隔阂，行业广泛转向采用外挂工具链的“智能体循环”（Agentic Loops）。但Flechter指出，智能体循环的自动化可靠性在实践中遭遇了明显的瓶颈，成功率普遍被限制在80%至85%之间。在严苛的企业生产安全场景中，15%至20%的失败率与幻觉容错率是完全不可接受的，这使得单纯依赖LLM构建的自主防御系统无法安全落地。
企业物理学与JEPA世界模型机制 (4:45)
针对大语言模型试错式提示词的脆弱性，Flechter提出必须建立“网络安全世界模型”（Cybersecurity World Model）。这一思路源自自动驾驶与机器人等物理实体AI（Physical AI）：自动驾驶车辆在驶入陌生十字路口时，依靠对速度、动量和转向等底层物理定律的理解完成转向，而非死记硬背历史文本数据。在数字领域，世界模型同样旨在将AI从语言翻译工具转变为懂得企业网络“底层物理学”的预测引擎。
在架构设计上，该模型借鉴了Meta首席AI科学家Yann LeCun倡导的联合嵌入预测架构（JEPA）及AMI等研究，摒弃了自回归逐词生成机制。该模型能够基于当前企业图谱状态X，在数字空间模拟配置变更（例如修改Amazon S3存储桶策略或调整权限规则），并准确预测变更后的未来状态，在规避试错操作与幻觉的同时，验证威胁是否被成功阻断且不破坏业务系统。
单次修复落地与遮蔽图训练方法 (9:12)
世界模型支持在无需人工核验幻觉的前提下实现安全环境的“单次自动修复”（Single-shot Remediation）。在云配置失误场景中，一旦出现开放的S3存储桶，传统排查需耗费大量精力调用日志与查询脚本以评估资产是否可穿透至核心数据库，而世界模型可在单一推理步中准确评估爆炸半径并给出阻断方案。同时，针对安全运维中“修复规则可能阻断正常微服务流量造成宕机”的普遍顾虑，世界模型能够在落地前预判图谱上的连通性变化，确保在不中断业务的前提下完成加固。
针对缺乏标准真值对照的训练与验证难题，Flechter在问答环节表示核心壁垒在于数据工程而非算力消耗。团队利用受JEPA启发的图遮蔽技术，构建了数千个包含用户、设备、权限和策略的高度结构化企业“虚拟世界”，通过遮盖部分子图与跃迁状态来训练模型推测完整图谱的能力。Flechter指出，安全控制本身即是数字世界的物理法则——如缺乏权限构成了绝对阻断屏障，缺失双因子认证（2FA）构成了绝对暴露路径，这些底层规则为世界模型的预测提供了坚实的数学依据。
Guy Flechter与Sola Security的核心结论是，通用语言模型的自回归机制因无法理解图谱物理学而无法独自胜任企业级网络安全防御，必须转向以JEPA等架构为核心的网络安全世界模型。然而，该路径的成败取决于能否高保真地合成极为复杂的异构企业环境拓扑，并在高速演化的企业IT架构与权限变更中持续保证预测数学真值的精确性。
Detailed notes — zh
## Background
该视频由 YouTube 频道 Ai4 于 2026 年 8 月 19 日发布，内容为 Sola Security 联合创始人兼 CEO Guy Flechter 的主题演讲。根据页面介绍，演讲深入探讨了大语言模型（LLM）在网络安全领域的固有局限性，并提出了克服这些瓶颈的潜在技术路径。
## TL;DR
Guy Flechter 指出企业网络安全本质上是由资产、权限和身份交织构成的拓扑图（Graph）问题，而擅长文本与列表处理的 LLM 在此面临根本瓶颈，其 80% 至 85% 的成功率无法满足自主防御系统的绝对可靠性要求。为打破这一僵局，Sola Security 借鉴自动驾驶与 Yann LeCun 提出的 JEPA 架构，倡导构建“网络安全世界模型”（Cybersecurity World Model）。通过将权限约束与安全控制定义为网络空间的“物理定律”，并在仿真世界中进行图掩码自监督训练，该模型能够直接预测系统状态转移并评估业务影响，从而真正实现零幻觉的单次推理与自主自愈防御。
## Atmosphere & style
演讲整体呈现出理性务实、逻辑严密的技术布道风格，主讲人 Guy Flechter 态度自信沉稳，开场以轻度幽默化解对前沿 LLM 巨头的技术挑战，随后迅速切入底层拓扑逻辑的深度拆解。现场互动与问答环节思想交锋激烈，直击测试数据合成与网络“物理法则”定义的底层难点，给观众带来一种穿透行业炒作迷雾、探索安全与 AI 交叉领域第一性原理的扎实沉浸感。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) 网络安全的本质是高度互联的“图”（Graph）问题，涵盖资产、身份、权限与服务的复杂拓扑，而非 LLM 擅长处理的线性语言与文本列表。
- [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) 防御者习惯用列表思考，而攻击者用图思考；只要 LLM 仍然用列表模式处理安全防御，攻击者就将持续占据优势。
- [01:55](https://www.youtube.com/watch?v=px24wf8UtPM&t=115s) 现有的 LLM 安全方案仅作为自然语言到 SQL、Cypher 等查询的翻译器，在复杂的企业真实上下文中成功率仅能达到 80%-85%，无法支撑绝对可信的自主防御系统。
- [04:39](https://www.youtube.com/watch?v=px24wf8UtPM&t=279s) 为了克服 LLM 的瓶颈，Sola Security 倡导引入借鉴自自动驾驶与机器人领域的“网络安全世界模型”，实现系统的单次状态转移预测。
- [06:06](https://www.youtube.com/watch?v=px24wf8UtPM&t=366s) 安全世界模型受 Yann LeCun 的 JEPA（联合嵌入预测架构）启发，让 AI 直接学习网络拓扑底层的“物理运行机制”，预测采取特定安全动作后的下一个系统状态。
- [09:12](https://www.youtube.com/watch?v=px24wf8UtPM&t=552s) 安全世界模型可实现“自主自愈”，在单次推理中量化攻击路径与阻断策略，摆脱昂贵的上下文工程与人工验证。
- [09:12](https://www.youtube.com/watch?v=px24wf8UtPM&t=552s) 该模型能够在拓扑图上直接预测安全配置变更对合法微服务流量的影响，避免传统修复引发业务宕机的高昂测试成本。
- [13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) 训练网络安全世界模型的最大挑战在于测试与仿真数据的构建，需要构建多维度的虚拟网络拓扑来严谨评估状态预测精度。
- [14:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=885s) 模型训练采用类似 JEPA 的图掩码机制，通过遮蔽图中的局部切片并让模型预测还原，强迫模型掌握网络环境内部的运行规律。
- [14:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=885s) 数据准备与拓扑建模是构建世界模型技术上最具挑战性的环节，而实际计算训练则是成本最高昂的阶段。
- [16:57](https://www.youtube.com/watch?v=px24wf8UtPM&t=1017s) 网络安全领域的“物理学定律”表现为硬性的安全控制约束（如权限缺失、缺少 2FA 等），这些确定性边界共同构成了安全世界的物理运行框架。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) 核心论点：网络安全本质是图问题而非语言列表
Guy Flechter 开宗明义地提出了 Sola Security 在过去两年中持续验证的核心技术论点：尽管大语言模型（LLM）能力强大（他打趣道“希望现场没有来自 OpenAI 和 Anthropic 的朋友”），但在网络安全领域存在着难以逾越的硬性瓶颈，无法真正实现规模化扩展。
为了阐明这一困境，Guy 引用了知名安全渗透工具 BloodHound 创始人的经典名言：“防御者用列表思考，攻击者用图思考；只要这种情况持续，攻击者就会获胜”（"defender thinks in lists, attacker thinks in graphs. As long as this is true, attackers will win."）。他强调，如果将这句话中的“防御者”替换为“LLM”，逻辑完全成立。LLM 本质上是在语言空间中运作，习惯将问题拆解为线性的“列表”并据此提供防护。然而，一旦脱离单纯的代码生成或代码审计（代码本质上属于语言问题），面对任何 20 人的初创团队或规模庞大的跨国企业，企业安全环境本质上全都是一个高度互联的“图”（Graph）——涵盖了资产（assets）、身份（identities）、权限（permissions）与服务（services）之间错综复杂的拓扑连接。
- **图与列表的错位**：LLM 试图以处理文本语言和离散列表的模式，去解决拓扑结构复杂的企业网络安全图问题。
- **环境的拓扑本质**：任何规模的企业网络都可以抽象并翻译为一张由资产、权限与身份互联构成的庞大关系图。
---
### [01:55](https://www.youtube.com/watch?v=px24wf8UtPM&t=115s) 现有 LLM 与智能体循环的瓶颈：85% 成功率不可接受
企业安全防御的终极目标或“圣杯”（holy grail），是构建一套能够接入企业环境、自主理解全局态势、具备推理能力并主动实施防护的全自动防御系统（fully autonomous defense system）。然而，当前市面上的 LLM 安全方案或 AI 原生平台，仅仅是将 LLM 当作自然语言的“翻译器”——负责将人类的自然语言提问翻译为 SQL、Python 或 Cypher 查询语句，再试图通过确定性的逻辑来给出答案。正如微软 CEO 所指出的，LLM 本身已逐步变成通用的基础设施商品（commodity），人人触手可及，而真正困难的是“上下文层”（context layer）。
向 LLM 解释企业的真实运行环境、判断资产权重极为艰难。即便耗费巨大精力构建了最顶尖的上下文层，现有基于 LLM 的方案在实际环境中的成功率也仅仅徘徊在 80% 至 85% 之间。业界曾寄希望于“智能体循环”（Agentic Loops）通过编排多种工具来弥补不足，但实践证明这依然无法彻底解决问题。对于追求绝对可靠的自主防御系统而言，85% 的成功率意味着存在不可忽视的安全漏洞与盲区，根本无法获得企业的真正信任。
```mermaid
flowchart TD
accTitle: LLM Limitations vs World Model Approach
subgraph CurrentApproach["当前 LLM 路线的局限"]
A["自然语言输入"] --> B["LLM 翻译层"]
B --> C["生成 SQL / Cypher / Python"]
C --> D["外部上下文层拼接"]
D --> E["执行确定性查询与编排"]
E --> F["80% - 85% 成功率 (存在幻觉与盲区)"]
end
subgraph WorldModelApproach["网络安全世界模型路线"]
G["企业全量资产与权限拓扑"] --> H["编码为统一企业图 (Graph)"]
H --> I["世界模型 (学习安全物理机制)"]
I --> J["单次状态转移预测 (Next-State Prediction)"]
J --> K["100% 自主自愈与可信防御"]
end
```
---
### [04:39](https://www.youtube.com/watch?v=px24wf8UtPM&t=279s) 破局之道：引入网络安全世界模型（World Model）
面对 LLM 方案的停滞不前，Sola Security 提出的演进方向是打造基础级的“网络安全世界模型”（Foundational Large Cybersecurity World Model）。Guy 向现场观众发问是否了解“世界模型”的概念，并解释该理念最初孕育自物理现实世界中的自动驾驶与机器人领域。
在物理世界中，世界模型的核心使命是帮助系统预测“当采取某项行动时，外部环境将会产生何种变化”。以自动驾驶汽车为例，车辆在路口右转时所面对的交通流和路况每一次都是全新的，但车载世界模型能够基于对物理规律的理解，精准预测在特定车速和转向角度下的未来状态。Guy 坚信，将世界模型引入网络安全领域，正是破解企业动态拓扑难题、迈向全自动防御系统的必由之路。
- **动态适应性**：与自动驾驶面对全新路况类似，企业网络中的每一次变更与威胁路径也各不相同，需要系统具备无须预先穷举的预测推理能力。
- **预测驱动而非规则驱动**：世界模型的核心在于理解状态演进机制，而非死记硬背过去的规则列表。
---
### [06:06](https://www.youtube.com/watch?v=px24wf8UtPM&t=366s) 从多查询生成到通用物理规律学习（JEPA 启发）
网络安全世界模型的核心，是用“原生世界智能”和“习得的物理机制”（learn physics）取代脆弱的多轮查询生成。AI 在这里不再充当文本翻译官，而是进化为一套安全领域的“物理引擎”。模型直接学习企业网络拓扑的底层运行逻辑，将安全分析从语言问题彻底拉回图状态预测问题。
Guy 提到了由图灵奖得主、Meta 首席 AI 科学家 Yann LeCun 提出的 JEPA（联合嵌入预测架构）以及前沿研究实验室 AMI。在这一架构下，世界模型通过学习通用的网络环境“物理定律”，专注于预测系统的“下一个状态”（next state）。例如，当系统处于状态 $X$ 时，若安全系统执行了操作 $Y$（如修改 S3 存储桶配置），模型能在不依赖反复生成查询、不产生大模型幻觉的前提下，直接推演出该变更将导致的系统状态以及它所能防御的攻击路径。
```mermaid
flowchart LR
accTitle: Next-State Prediction in Cybersecurity World Model
subgraph PredictionCycle["世界模型状态预测闭环"]
S0["当前状态 X (State X)"] --> Act["拟执行安全动作 (Action Y/Z)"]
Act --> WM["世界模型 (物理引擎推演)"]
WM --> S1["预测未来状态 (State X')"]
S1 --> Dec["评估阻断效果 / 业务影响"]
end
```
---
### [09:12](https://www.youtube.com/watch?v=px24wf8UtPM&t=552s) 核心应用场景：自主自愈与零影响修复预测
一旦安全世界模型落地，将解锁极为广阔的应用潜力，其中最具革命性的当属“自主自愈”（autonomous self-healing）。防御系统可无缝接入企业环境，自动将分散的身份、资产与权限关系编码为图模型，在完全不需要人工介入验证幻觉（no human in the loop）、无需繁重外部上下文工程的前提下，自主推演修复操作的后果并实施闭环防护。
Guy 在现场分享了两个具有代表性的实战场景：
1. **云配置失误与攻击路径秒级量化**：当检测到 S3 存储桶公开或安全控制缺失时，世界模型能够在“单次推理”（single shot）中直接量化攻击者是否能够借此路径横向渗透至生产核心数据库，并快速给出最优拦截策略，告别繁琐的上下文检索与多轮提问。
2. **修复措施的业务影响评估（防宕机）**：安全团队在实施配置变更时最大的痛点在于“该操作是否会误伤合法的微服务流量并引发业务中断”。传统流程依赖漫长的日志比对、沙盒验证与人工测试；而世界模型则能直接在图结构上预测配置变更后的流量状态变化，以极低的时间、算力与资金成本完成无害化评估。
---
### [13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) 问答环节：世界模型的评估与测试数据挑战
在演讲后的问答环节中，现场观众提问：既然世界模型的核心是预测未知的未来状态，且在现实环境中往往难以立即与“标准答案”进行比对，那么研发此类模型时面临的最大测试与精度验证挑战是什么？
Guy 回答表示，最核心的瓶颈在于**测试数据的构建**。训练与验证世界模型完全不同于传统 LLM——后者只需从互联网抓取数十亿文本数据并调整模型权重；而构建网络安全世界模型，则必须在底层合成并模拟出各种不同拓扑、不同业务逻辑与攻击路径的“虚拟世界”（synthetic worlds），唯有通过多维度的世界模拟，才能严谨量化并验证模型对未来状态的预测精度。
### [14:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=885s) 基于 JEPA 与图掩码的世界模型训练方法及数据准备挑战
从高维度来看，训练这类网络安全世界模型的核心机制类似于创建大量仿真“世界”。系统向模型展示这些世界结构，并在生成新世界时故意遮蔽图结构（Graph）中的某个局部切片，随后要求模型去预测或填补被隐藏的部分，以此掌握环境内部的潜在关联与运行规律。这种训练思路很大程度上借鉴了 Yann LeCun 及其团队提出的 JEPA（Joint Embedding Predictive Architecture，联合嵌入预测架构）概念。虽然业界存在多种不同类型的世界模型，但底层的核心都在于对数据结构的表征与预测。
在数据构建方面，模型首先需要吸纳来自多样化环境的海量配置数据。然而，单纯积累配置只是第一步，真正的难点在于如何组织并结构化这些数据——如何将碎片化的数据点相互连结，精准描摹出一个组织内部用户（User）、设备（Device）、服务（Service）、权限（Permission）、具体行为活动（Activity）以及安全策略（Policy）之间的复杂网状关系。
模型训练过程清晰地划分为两个阶段：数据准备阶段与实际训练阶段。实际的计算训练固然耗资巨大、极为昂贵，但前期的数据准备与拓扑建模才是技术上最具挑战性的环节。这也是为何大语言模型（LLM）厂商往往热衷于宣称模型经过了 300 亿、400 亿甚至更多数据点的训练，其背后的潜台词正是攻克海量且高质量训练数据集构建这一最大难关。
```mermaid
flowchart TD
accTitle: 世界模型图拓扑训练流程
subgraph S1["数据准备阶段（最核心挑战）"]
A["多环境配置数据"] --> B["构建实体关联图谱<br/>(用户 - 设备 - 服务 - 权限 - 策略)"]
end
subgraph S2["JEPA 掩码自监督训练"]
B --> C["生成仿真网络世界拓扑"]
C --> D["遮蔽局部图结构 (Masking)"]
D --> E["世界模型预测与还原补全"]
end
subgraph S3["实际算力训练（最高成本）"]
E --> F["更新世界模型表征权重"]
end
```
- **训练机制核心**：基于 JEPA 架构，通过在网络拓扑图中随机遮蔽部分节点或连接，强迫模型预测缺失的图切片，从而学会安全环境的内部推演规则。
- **实体关联图谱维度**：包含用户（User）、设备（Device）、服务（Service）、权限（Permission）、操作活动（Activity）与安全策略（Policy）等多维联动。
- **两阶段难点对比**：数据准备是智力与技术层面“最具有挑战性”的部分，而模型训练运算则是资金与算力层面“最昂贵”的部分。
---
### [16:57](https://www.youtube.com/watch?v=px24wf8UtPM&t=1017s) 现场问答：网络安全领域的“物理学定律”与底层规则
在现场互动环节，观众提出了一个深刻的本质问题：在物理世界模型中，存在麦克斯韦方程组、牛顿力学定律等确定性的物理真理作为客观基准（Ground Truth）；但网络空间的行为表现极其多变且动态，那么网络安全世界的“方程”或“物理学定律”到底是什么？模型该如何计算并预测网络安全世界的行为？
Guy Flechter 回应称，这正是 Sola Security 技术中的“秘密配方”（Secret Sauce）。他们曾针对“AI 原生资产”（AI Native Assets）发表过专门的论文，阐述如何构建这类动态拓扑图及其内部关系。在网络安全领域，所谓的“物理定律”实际上体现在底层的确定性安全规则与控制约束之中。例如，“权限缺失”（Lack of Permission）就是一个不可逾越的客观事实和铁律，缺乏双因子认证（Lack of 2FA）同样构成了某种环境或情境下的安全物理边界。
通过在资产关联图谱之上叠加安全控制措施（Security Controls），并精准定义特定情境下“应当发生什么”与“绝对禁止发生什么”，便构成了网络安全世界的物理运行框架。Guy Flechter 坦言，与物理世界中根据视觉直观理解椅子形态或判断行进方向相比，将复杂的安全交互形式化翻译为规则确实并非易事，但这完全是可实现的，且团队正沿着这条技术路径持续推进。
- **物理世界的确定性 vs 网络空间**：物理世界依赖经典力学或电磁学方程，而网络安全世界模型则依赖资产拓扑图及确定性的安全约束规则作为底层基准。
- **安全的“物理法则”实例**：权限的有无、双因子认证的缺失等硬性访问边界，在拓扑逻辑中即充当了安全世界的确定性物理规则。
- **拓扑与控制叠加**：在 AI 原生资产拓扑之上映射安全控制策略，界定系统在特定场景下允许与禁止的行为边界。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) defender thinks in lists, attacker thinks in graphs. As long as this is true, attackers will win. — BloodHound Founder (cited by Guy Flechter)
> [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) 网络安全领域存在着难以逾越的硬性瓶颈，无法真正实现规模化扩展。 — Guy Flechter
> [00:04](https://www.youtube.com/watch?v=px24wf8UtPM&t=4s) 企业安全环境本质上全都是一个高度互联的“图”（Graph）——涵盖了资产、身份、权限与服务之间错综复杂的拓扑连接。 — Guy Flechter
> [01:55](https://www.youtube.com/watch?v=px24wf8UtPM&t=115s) LLM 本身已逐步变成通用的基础设施商品，人人触手可及，而真正困难的是“上下文层”。 — Guy Flechter
> [01:55](https://www.youtube.com/watch?v=px24wf8UtPM&t=115s) 对于追求绝对可靠的自主防御系统而言，85% 的成功率意味着存在不可忽视的安全漏洞与盲区，根本无法获得企业的真正信任。 — Guy Flechter
> [04:39](https://www.youtube.com/watch?v=px24wf8UtPM&t=279s) 在物理世界中，世界模型的核心使命是帮助系统预测“当采取某项行动时，外部环境将会产生何种变化”。 — Guy Flechter
> [06:06](https://www.youtube.com/watch?v=px24wf8UtPM&t=366s) 世界模型通过学习通用的网络环境“物理定律”，专注于预测系统的“下一个状态”。 — Guy Flechter
> [13:49](https://www.youtube.com/watch?v=px24wf8UtPM&t=829s) 构建网络安全世界模型，则必须在底层合成并模拟出各种不同拓扑、不同业务逻辑与攻击路径的“虚拟世界”。 — Guy Flechter
> [14:45](https://www.youtube.com/watch?v=px24wf8UtPM&t=885s) 数据准备是智力与技术层面“最具有挑战性”的部分，而模型训练运算则是资金与算力层面“最昂贵”的部分。 — Guy Flechter
> [16:57](https://www.youtube.com/watch?v=px24wf8UtPM&t=1017s) 在网络安全领域，所谓的“物理定律”实际上体现在底层的确定性安全规则与控制约束之中。 — Guy Flechter
## Mentioned
Guy Flechter, Sola Security, BloodHound, OpenAI, Anthropic, Microsoft, Yann LeCun, JEPA (Joint Embedding Predictive Architecture), AMI, Meta, AWS S3, Cypher, SQL, Python
## Video description
Guy Flechter - Co-Founder and CEO, Sola Security
We will discuss the limitations of LLM for Cybersecurity and what possible ways to overcome them.
[Source](https://www.youtube.com/watch?v=px24wf8UtPM)
Source quotes
[0:04] BloodHound Founder (cited by Guy Flechter): defender thinks in lists, attacker thinks in graphs. As long as this is true, attackers will win.
[0:04] Guy Flechter: 网络安全领域存在着难以逾越的硬性瓶颈，无法真正实现规模化扩展。
[0:04] Guy Flechter: 企业安全环境本质上全都是一个高度互联的“图”（Graph）——涵盖了资产、身份、权限与服务之间错综复杂的拓扑连接。
[1:55] Guy Flechter: LLM 本身已逐步变成通用的基础设施商品，人人触手可及，而真正困难的是“上下文层”。
[1:55] Guy Flechter: 对于追求绝对可靠的自主防御系统而言，85% 的成功率意味着存在不可忽视的安全漏洞与盲区，根本无法获得企业的真正信任。
[4:39] Guy Flechter: 在物理世界中，世界模型的核心使命是帮助系统预测“当采取某项行动时，外部环境将会产生何种变化”。
[6:06] Guy Flechter: 世界模型通过学习通用的网络环境“物理定律”，专注于预测系统的“下一个状态”。
[13:49] Guy Flechter: 构建网络安全世界模型，则必须在底层合成并模拟出各种不同拓扑、不同业务逻辑与攻击路径的“虚拟世界”。
[14:45] Guy Flechter: 数据准备是智力与技术层面“最具有挑战性”的部分，而模型训练运算则是资金与算力层面“最昂贵”的部分。
[16:57] Guy Flechter: 在网络安全领域，所谓的“物理定律”实际上体现在底层的确定性安全规则与控制约束之中。
Source quotes
[0:04] Guy Flechter: defender thinks in lists, attacker thinks in graphs. As long as this is true, attackers will win.
[13:49] Guy Flechter: It's not like taking LLM bringing billions of data points or scraping the internet and start putting weights inside the model. You need to create a lot of types of worlds.
Original transcript
[0:04] Uh so, although I have only 20 minutes, I will
try a little bit uh to explain why we think it's
[0:12] Solar again, it's our thesis in the last
2 years, that's what we're working, how
[0:18] and why we think that LLMs are, which are very
good, I hope no OpenAI and Anthropic in the crowd,
[0:24] uh will not scale. Oh, they have hard limitation
about what they can do in cybersecurity.
[0:32] I would like to start with a phrase that
I'm not sure if everybody's familiar with.
[0:36] Uh this is something that from uh the the the
person that uh innovated BloodHound, if anybody
[0:44] knows, this is a security tool that helped to
understand path inside uh your organization,
[0:51] and it says uh defender thinks in lists, attacker
thinks in graphs. As long as this is true,
[0:55] attackers will win. Uh you can replace defenders
with LLMs, and this is the same thing, this is
[1:01] the same problem. LLMs are thinking in language
problem, thinking about what they can do in lists,
[1:07] and and then they try to again, uh uh uh
protect you, but when we're thinking about
[1:15] uh uh cybersecurity outside of code, okay, let's
go outside of the code area, which again, code is
[1:23] a language problem or a language situation. When
we go outside of uh the code and talking about uh
[1:31] enterprise, taking any organization whatsoever,
whether it's a small organization of 20 employees,
[1:36] and when you go out to large
organization, take any big organization,
[1:41] it's a graph problem, it's a relationship between
uh assets and identities, permissions, services,
[1:49] and so on, and everything is connected, and
you can again translate it into a graph.
[1:55] And and when we think about what is the end goal
or what is the holy grail of what we're trying
[2:03] to achieve is fully autonomous defense system
which are can connect into your environment,
[2:09] understand by themselves, and then act, resonate,
and and protect you for everything. But uh uh
[2:18] today uh what are doing today the LLM solution or
or AI native platform use LLM to translate. Okay,
[2:29] translate natural language. Okay, or they try
to translate natural language into queries, SQL,
[2:37] Python, or Cypher. And then try to move it into
a deterministic approach, and then try to explain
[2:45] on answer something, but they are unable to do it
by themselves. They need something in between them
[2:53] to the environment. This is why when everybody
is speaking about it, everybody is speaking that
[2:59] the LLM or for example the Microsoft CEO uh speak
about that LLMs are commodity. Everyone can reach
[3:08] anyone in this room, anyone outside can reach a
LLM, but the hardest part is the context layer.
[3:13] Explaining the LLM the environment, explaining
them what is how the environment operates,
[3:20] what is important, what is not important, and
they are struggling, and this is why we see that
[3:28] even with the best uh uh context layer, they
are getting into 80, 85% uh success rate.
[3:40] Uh so we thought that Agentic Loops
can uh uh solve it, and they didn't.
[3:48] Uh at the end of the day, like I said, uh we are
orchestrating uh tools and explain to them and try
[3:56] to explain to them uh what's next and giving them
uh deterministic um understanding of environment,
[4:05] and this is not working. And we see it and we're
struggling with it. And like I said, even the best
[4:12] L L O A I native platform are getting into 80
85% success rate and going back to a holy grail
[4:20] about autonomous defense system, we cannot
trust a fully autonomous defense system with
[4:26] 85% uh success rate. Uh it's not something
that can be achieved or can be uh trusted in
[4:33] what we think can be the future. Uh so, this is
uh a little bit about the problem or where we
[4:39] are standing around the problem. And what we are
thinking about is that the future it should be
[4:45] a fundamental building the foundational large
cybersecurity world model. Anybody familiar with
[4:51] the concept world model? Raise your hand if you're
familiar with a world model. Okay, a world model
[4:59] is something that uh is coming from the physical
world. When you think about autonomous uh cars,
[5:06] uh robotics, and stuff like that, a world model
help uh such areas to interact or predict what
[5:17] will happen in case that something or they need
to do something. For example, take autonomous car,
[5:23] they need to take turn right, and it should be
exactly as good as the previous ride that they
[5:31] took. Uh and it's and it's a new ride. Okay?
It's completely new ride. Every ride that a car
[5:37] autonomous car is taking, it's completely new and
completely different from the previous ride. So,
[5:44] uh uh this is a little bit about world model
and what it's being used to. And we think that
[5:50] a world model, this is the future for cyber
security and this is what we believe should be
[5:56] the the path for the holy grail of building
and fully autonomous defense system.
[6:06] Okay. So, what we are replacing is replacing or
giving native world intelligence replacing brittle
[6:15] queries with learn physics. Actually, we try to
tell the foundational model again, it's replacing
[6:23] LLM. It's no longer a language problem. This is
This is what we're saying. It's not a language
[6:29] problem. It's a graph problem and a world
model should be able to do it. And there are
[6:35] different types of world models, the training
of a world model, for example, Jepa, which was
[6:42] invented but by Jan LeCun, uh that was in charge
of the AI in Meta super intelligence and now is
[6:50] the chairman of AMI, it's frontier lab for world
models. Uh so, like we said, it's replacing
[6:59] the multi-query generation into universal
world learning which means is learning
[7:07] the physics, okay? And a model that in that kind
of a model it at the end of the day try to predict
[7:16] what is the next state, okay? Which means that,
okay, we have state X. If we would do X, Y, or Z,
[7:23] what would be the outcome of that? If we need
to replace a a configuration in S3 bucket,
[7:32] what would be the outcome or what we will prevent
in the next state of the change. It's not try to
[7:40] do a lot of queries or try a lot of things and
then there is hallucination and all the other
[7:46] stuff. It's predicting. It's predicting what
is the outcome of actions and what it will
[7:52] be the state after that kind of action. Again,
going back to for example, like autonomous car,
[7:59] if I need to take the turn right, what will be
the outcome of doing that turn in that kind of
[8:05] velocity, in that kind of uh uh uh shifting
the the wheel in that angle and and so on.
[8:13] And this is exactly what a world model
for uh uh uh for cybersecurity will do.
[8:23] So, what is again, going back about it's
and no longer AIs is no longer a translator,
[8:32] it's a physics engines. Again, it's explaining the
world, it's connecting and looking on environment
[8:39] and looking on connection of identities,
assets, permission, then each type of company,
[8:46] it's a new company and then they try to translate
it and encode the enterprise into a graph and then
[8:54] try to say, "Okay, what will be the
future state when I will do X or Y in that
[9:02] uh environment?" And that can lead to the
holy grail of fully autonomous defense system.
[9:12] Don't worry, I will save some time for more
question at the end. So, those are some of the
[9:19] core value proposition and use cases that can
be done, but like we said, there are enormous
[9:27] things that can be done with the world model
once it will be available, but the best
[9:33] one of the most important thing is the autonomous
self-healing, which means that a system can
[9:39] connect to your environment, spread across
everything that you have inside your environment,
[9:45] and then start to do a self-healing, and start to
detect, change, and then protect your environment
[9:53] by understanding what will be the outcome of
doing so without, again, the need of any context
[10:01] of your environment. It will understand by itself
without the need of any human in the loop of
[10:10] checking if there is hallucination of doing
so, and anything like that. That's what the
[10:17] the LLM system with a world
model will be able to do in that.
[10:25] And some uh uh scenarios, uh for example,
a cloud misconfiguration is detected.
[10:34] The outcome or the goal is to instantly
quantify if an attacker can reach production
[10:40] database. For example, if for S3 bucket uh
is open, or there is a permission that uh is
[10:48] uh lacking of specific security control, we want
to understand very very fast in a single shot
[10:55] if we would do X, or what it what we should
do in order for us to prevent from an attacker
[11:03] to get to a certain database, or a very important
asset inside a production. So, today, the current
[11:11] approach is that I need to bring a full context,
try to explain the LLM what it need to do,
[11:18] and resonate, and then go into deterministic
approach. But, in a world model solution,
[11:26] it's doing it by itself, understanding all the
all the possibilities, predicting what will be
[11:34] the best outcome, and then just doing that and
move on. That's That's the difference. There is no
[11:42] uh back and forth of understanding, or there is
a lot of explaining to the LLM what to do next.
[11:50] In a world model, already predicting
and then doing it instantly.
[11:59] Another scenario, uh again, uh
configuration remediation effect, uh
[12:08] uh what we could The critical question,
will this change block legitimate
[12:13] microservice traffic and cause an
outage? Which means, if I would do X,
[12:18] will will I'm doing any damage to my company? Will
I do anything that prevent employees to do so?
[12:27] And again, what we do up until now, log deduction,
checking, testing, verifying, and so on.
[12:38] In the future, uh with a world model, easy. We
are predicting the graph of building a graph
[12:45] of what can be the outcome, or what is the outcome
of changing something in the logs, or changing a
[12:51] service, or changing configuration of a service,
and then just doing that because we understand
[12:58] what is the prediction, or what is the next
state of the environment what after I will do
[13:06] something that change this configuration. Those
are just two types of uh uh uh examples, but
[13:15] again, think of any type of change, any type of
defense, any type of a use case that can be or
[13:23] need to be done in in day-to-day, uh if you are
in the security team and a world model should be
[13:30] able to do it in a very easy way in a single shot
and very very very important very efficient from
[13:38] time money and and resources that need to be
invest in order for the the world model to do it.
[13:49] That's it. Very quick. Thank you. Questions? Guy,
can you maybe pause for a second and explain how
[13:58] do you or what are the challenges in testing such
models or showing accuracy level because as you
[14:05] mentioned those models meant for predicting the
future basically and not it's very hard I can only
[14:12] assume to test those models right now but when you
don't know the outcome and you can't compare it
[14:17] to the to the preferred outcome so how about the
challenges? Yeah, so the the good question so the
[14:24] challenges at the end of the day data. Creating
the right type of data in order for you to test
[14:30] it. It's not like taking LLM bringing billions
of data points or scraping the internet and start
[14:39] putting weights inside the model. You
need to create a lot of types of worlds.
[14:45] Basically in high level how you train that
kind of a model it's you create lot of worlds
[14:52] explained the model look on those what create
a new world hide piece of the graph of the new
[15:00] world and then tell the the model okay try to
guess what should be this part in the graph
[15:07] in order for you to to predict what need to be
there. This is like in a very high level how
[15:14] how you train that kind of world. Again
it's a base on the concept of a Japa.
[15:20] It's a it's type of training or the world
model that Jan and his team invented. Uh,
[15:27] there are many types of world models that
like I said. Um, so yeah, data first of all,
[15:35] uh, you need to have configuration data, a lot of
again, type of configuration data from different
[15:41] environment, but also how you structure the data,
how you connect this type of data to each other,
[15:47] how you build and explain organization, the
connection between uh, user to device to service
[15:55] to permission and all everything else and activity
and policies and and so on. So, it's not uh,
[16:04] uh, that easy, uh, but there is a very uh, uh,
clear path of how to do it and this is the biggest
[16:13] challenge by the way in there are two type of the
the two parts of training the model. There is the
[16:20] preparation of the data and there is the training
itself. The training itself is the expensive part.
[16:25] Uh, this what cost the most, but the preparation
of the data that that's the most challenging part.
[16:33] Uh, this is why when we're looking on
LLM, the the main thing that they always
[16:37] try to say that we train it on 30 billion
data points or 5 billion or 40 billion.
[16:42] That's that's what they're trying to explain that
the the biggest challenge is how big the data set
[16:49] they managed to uh, prepare for the model to
be trained on. Oh, we have another question.
[16:57] Hi. Um, my question is when you talk
about when you think about a world models,
[17:03] physical world models, um, you have ground
truth which is the physics, Maxwell equations,
[17:09] Newton equation and so on. What are the equations
for cyber? I mean, networks behave differently
[17:16] and very dynamically. So, what will be the the
equation? How do you know how the world do to
[17:22] calculate how the world of cyber behaves?
So, this is a very good question. This is uh
[17:29] this is a little bit the secret sauce. Uh how
you create the graph and the relationship and
[17:35] and and and create we publish a paper around
AI native assets, how you create this kind of
[17:43] topology. Uh but also on top of that, there is a
lot of cyber research on that domain to explain
[17:51] the physics to get to the physics again. Lack
of permission, it's lack of permission. This is
[17:58] physics. This is the rule of lack of two-factor
authentication. It's It's You can say it's a
[18:05] like uh explain uh physics uh of uh of environment
or situation. Um so, on top of how you build uh
[18:16] again, try to simplify it. On top of how you
build the relationship between the asset,
[18:21] you put on top of that how you are putting
a security control and again, explaining the
[18:27] physics of what should be done or not to be done
in that kind of situation. Again, it's not that
[18:34] uh easy to translate it because uh in physical
world, it's very easy to explain. You take the
[18:40] right how you should take the right or vision,
what is chair look like and and and so on. But
[18:48] uh we believe that's uh doable
and on and on the path to do that.
