Visa 将其 AI 驱动的网络防御系统开源
英文标题：🔶 Visa Releases Its AI-Powered Cyber Defense System to Open Source
Visa 宣布将 Visa Vulnerability Agentic Harness（Visa 漏洞 Agentic Harness）开源，这是一套可在大规模代码与系统环境中部署 AI Agent、发现漏洞、排序优先级并生成修复报告的安全测试工具。Visa 表示，该工具已扩展至支持漏洞修复与补丁验证，但关键决策仍由人工安全和工程团队监督。作为 Anthropic Project Glasswing 的参与者，Visa 称该项目首月已在广泛使用的关键软件中发现超过 1 万个高危或严重漏洞。Visa 的核心结论是，未来网络防御的关键不再只是更快发现漏洞，而是以相近速度完成验证、修复和闭环确认。
65.7 KB
90.5 KB
100.1 KB
中文
随着人工智能加速发展，网络威胁也正以同样快的速度演变。对于防御者而言，这要求具备速度、纪律性，并持续聚焦于韧性。
为何重要
Visa 每年处理数千亿美元规模的交易。多年来，我们通过零信任架构、分层防御，以及为满足全球支付所需规模和可靠性而构建的自动化安全运营体系与平台，不断加固我们的网络。我们加入了 Anthropic 的防御性网络安全计划 Project Glasswing，以 AI 的速度测试这些防御措施，并更好地了解先进 AI 能如何帮助我们进一步提升能力。
Anthropic 的初步更新有助于量化这一变化的规模。Project Glasswing 参与者在测试的第一个月内，已在广泛使用、具有系统重要性的软件中识别出超过 10,000 个高危或严重级别漏洞。Anthropic 还表示，软件安全的进展如今已不再受限于发现漏洞的速度，而是受限于验证、披露和修复漏洞的速度。¹
以 AI 的速度测试我们的防御能力
为支持这些工作，我们的团队开发了一套多模型安全测试套件（现已迭代至第五代）。该套件围绕先进的扫描框架构建，可系统性地映射我们的代码库，在整个环境中部署 AI Agent 以发现漏洞，对发现结果进行分类和优先级排序，并为开发人员及修复 Agent 生成详细报告。我们正通过开源方式提供 Visa Vulnerability Agentic Harness，以推动更广泛社区中的防御创新。
我们构建 Visa Vulnerability Agentic Harness，是为了让其以 AI 的速度运行；我们也将以同样的方式持续演进它。自将其开源以来，我们已扩展其能力，使其不仅能帮助团队识别问题，还能快速、安全地支持修复和验证。AI 正让代码缺陷更容易被发现，但更棘手的问题是下一步该怎么做。通过这些更新，Visa 已将修复和验证纳入工作流，从识别哪些地方出现问题，进一步转向帮助团队解决这些问题。
在工作流的每一个关键节点，人工监督仍然不可或缺。我们的安全和工程团队会审查发现结果、验证严重程度，并在推进变更前确定修复路径。补丁验证 Agent 随后会自主处理并验证修复结果，从而简化流程、提高可靠性，以加速专家决策，而非取代专家决策。
该测试套件的最新一代能够理解自然语言目标，适应我们系统中的变化，并在无需手动维护脚本的情况下持续运行。这使我们的团队能够以与当前威胁发现速度相匹配的节奏开展工作，同时不牺牲保障网络韧性所必需的人类判断。
目前的三项经验
系统规模下的高保真分析 —— 该模型展现了执行全系统、意图感知和上下文感知分析的能力，产出的发现结果保真度高、噪声低且可直接采取行动。Mythos 在识别技术栈深层漏洞，以及突出显示在相互串联后可能变得更严重的问题方面表现尤为突出。
平均适应时间比检测时间更重要 —— 发现漏洞已不再是最困难的部分。真正的挑战在于，一个团队能够多快确认某个问题确实可被利用、修复该问题，并证明攻击路径已被关闭，而不只是证明补丁已被应用。我们将此称为“平均适应时间”（Mean Time to Adapt）。我们通过以下指标进行追踪：资产清单的时效性、每次发布后仍然存在的真实攻击路径数量，以及我们能够多快证明某项修复已在生产环境中生效。随着检测速度加快，以相近速度完成修复和验证的能力，将日益成为有效网络防御的决定性因素。
供应链风险正在上升 —— 在 AI 加速的威胁环境中，供应链风险变得更加重要。开源软件和第三方软件中的漏洞，可能迅速转化为实际暴露面。Visa 正积极参与行业行动，例如 IBM 和 Red Hat 的 Project Lightwell，通过 AI 驱动的验证和协同补丁修复，在大规模环境中强化开源安全。同时，我们也正更加聚焦供应商安全，利用漏洞指标和可利用性洞察来识别风险最高的环节，并确保供应商达到我们技术栈所要求的韧性标准。
Mythos 证实了什么
Mythos 在 Visa 发现了什么？我们的防御体系经受住了考验。一些发现结果被标记为严重问题，但我们的零信任控制、网络分段和现有保障措施本可阻止漏洞被利用。换言之，在外部行为者能够利用这些漏洞之前，攻击链就已经被打断。它发现了新的漏洞，也展示了我们可以改进的方向，同时还确认我们的安全基础——包括零信任架构——依然有效，并能够适应当前挑战。
下一步计划
在 Mythos 之后的世界里，其影响不只是更快地发现问题。随着 AI 压缩缺陷创建、发现、构建利用程序和实际投入使用之间的时间，网络防御将需要从围绕告警和人工分诊构建的模式，转向能够跨信号推理、应用策略护栏，并以机器速度采取有限行动的架构。这包括在明确治理下，持续推动组织同时规模化检测与修复能力。
此外，Mythos 并非孤例。其他 Frontier 模型和快速进步的替代方案正在迅速缩小差距，这表明此类能力将广泛可及。
Visa 的应对措施围绕三项优先事项展开：
- 通过在代码和构建流水线中进一步左移安全能力来缩小攻击面，使可被利用的漏洞在进入生产环境之前就被设计消除。
- 在高风险、缺乏维护支持的开源和商业组件演变为重大风险暴露之前替换它们，从而消除对供应链的结构性依赖。
- 重构我们的防御体系，使其日益自主化，让检测、验证和响应能够随威胁数量及零日漏洞复杂度同步扩展，同时始终处于清晰的人工治理之下。
这些优先事项体现了 Visa 对防御性 AI 未来的看法：它不只是更好地发现问题的方法，更是能够在由治理和人工监督界定的框架内，强化验证、优先级排序和响应的方式。未来的攻击者将不会像人类一样思考、行动或扩展规模，因此，建立在特征签名、静态态势模型和被动人工响应基础上的传统保护机制已不再足够。我们正通过设计面向 AI Agent 的控制平面来应对这一挑战；该控制平面专为以 AI 的速度实现自主实时检测、防御、修复和遏制而设计，在每一层内置韧性，并在每一个关键决策点保留人工问责。
这一转变强化了长期以来指导 Visa 安全方法的一项原则：韧性必须内建于每一层。我们的多层安全模型旨在实时检测、预防、响应和恢复应对威胁，同时维持网络的可靠性和完整性。随着威胁格局不断演变，我们正部署和评估先进的防御性 AI 能力，以进一步提升安全态势，并将这些经验用于加强我们所服务生态系统中的保护措施。
攻击者已经在使用这些工具。那些及早以这种纪律性部署防御性 AI 的组织，将成为能够守住防线的组织。对 Visa 而言，目标简单且始终如一：保障全球资金流动的安全，服务所有依赖它的人。
如需深入了解 Visa 从 Project Glasswing 中获得的经验，以及这对网络韧性意味着什么，请查阅下方完整技术简报。
English / 原文
As artificial intelligence accelerates, cyber threats are evolving just as quickly. For defenders, that requires speed, discipline, and a continued focus on resilience.
Why this matters
Visa moves hundreds of billions of transactions annually. Our network has been hardened over many years through zero trust architecture, layered defenses, and automated security operations and platforms built for the scale and reliability that global payments demand. We joined Project Glasswing, Anthropicâs defensive cybersecurity initiative, to test those defenses at AI speed and better understand where advanced AI can help us go further.
Anthropicâs initial update helps quantify the scale of the shift. Project Glasswing participants identified more than 10,000 high- or critical-severity vulnerabilities across widely used, systemically important software in the first month of testing. Anthropic also said that progress in software security is no longer limited by how quickly vulnerabilities can be found, but by how quickly they can be verified, disclosed, and patched.¹
Testing our defenses at AI speed
To support these efforts, our team developed a multi-model security test suite (now in its fifth generation) built around an advanced scanning harness that systematically maps our codebase, deploys AI agents across our environment to uncover vulnerabilities, categorizes and prioritizes findings, and generates detailed reports for developers and remediation agents. We are making the Visa Vulnerability Agentic Harness available through open source to help advance defensive innovation across the broader community.
Weâve built the Visa Vulnerability Agentic Harness to operate at AI speed, and we will continue to evolve it in the same way. Since open sourcing it, weâve expanded its ability to help teams not just identify issues but support remediation and validation quickly and safely. AI is making it easier to find flaws in code, but the harder question is what to do next. With these updates, Visa has added remediation and validation to the workflow to move beyond identifying what is broken and toward helping teams fix it.
Human oversight remains essential at every key point in the workflow. Our security and engineering teams review findings, validate severity, and determine remediation paths before changes are advanced. Patch validation agents then address and validate fixes autonomously, streamlining the process and improving reliability to accelerate expert decision-making, not replace it.
The latest generation of the suite can interpret plain-language objectives, adapt to changes across our systems, and operate continuously without manual script maintenance. This allows our teams to operate at a tempo that matches how quickly threats are now discovered, without sacrificing the human judgment that keeps our network resilient.
Three lessons so far
High-fidelity analysis at system scale â The model demonstrated the ability to perform system-wide, intent-aware, and context-aware analysis, producing findings that were high fidelity, low noise, and actionable. Mythos showed particular strength in identifying vulnerabilities deep in the stack and highlighting issues that may become more serious when chained together.
Mean Time to Adapt matters more than time to detect â Finding vulnerabilities is no longer the hard part. The real challenge is how quickly a team can confirm that an issue is truly exploitable, fix it, and then prove the attack path is closed rather than simply showing that a patch was applied. We call this Mean Time to Adapt. We track it by how current our inventory is, how many real attack paths remain after each release, and how quickly we can show that a fix works in production. As detection accelerates, the ability to remediate and validate at similar speed will increasingly define effective cyber defense.
Supply chain risk is rising â Supply chain risk matters even more in an AI-accelerated threat environment. Vulnerabilities in open-source and third-party software can quickly become real exposure. Visa is actively engaging in industry efforts such as IBM and Red Hatâs Project Lightwell to strengthen open-source security at scale through AI-driven validation and coordinated patching. In parallel, we are sharpening our focus on vendor security, using vulnerability metrics and exploitability insights to identify where risk is highest and ensure vendors meet the resilience standards required for our technology stack.
What Mythos confirmed
What did Mythos show at Visa? Our defenses held. Some findings were flagged as critical, but our zero trust controls, network segmentation, and existing safeguards would have prevented exploitation. In other words, the kill chain was broken before an external actor could act on those vulnerabilities. It found new vulnerabilities and showed where we can improve, while also confirming that our security foundations, including our zero trust architecture, remain effective and appropriate to the challenge.
What comes next
In a post-Mythos world, the implications go beyond faster discovery. As AI compresses the time between defect creation, discovery, exploit construction, and operationalization, cyber defense will need to evolve from models built around alerts and human triage toward architectures that can reason across signals, apply policy guardrails, and take bounded action at machine speed. This includes continued progress in how organizations scale both detection and remediation together under clear governance.
Furthermore, Mythos is not an outlier. Other Frontier models and rapidly advancing alternatives are quickly closing the gap, signaling that these capabilities will become broadly accessible.
Visaâs response is organized around three priorities:
- Reduce the attack surface by shifting security further left in our code and build pipelines, so exploitable vulnerabilities are designed out before they reach production.
- Remove structural dependency on the supply chain by replacing high risk, under supported open source and commercial components before they turn into material exposure.
- Refactor our defenses to be increasingly autonomous so detection, validation, and response can scale with threat volume and zero day sophistication, while staying under clear human governance.
These priorities reflect how Visa sees the future of defensive AI: not simply as a better way to find issues, but as a way to strengthen validation, prioritization, and response within a framework defined by governance and human oversight. The future attacker will not think, operate, or scale like a human, which is why traditional protection mechanisms built on signatures, static posture models, and reactive human response are no longer sufficient. We are addressing this challenge by designing a control plane for AI agents, engineered for autonomous real time detection, defense, remediation and containment at AI speed, with resilience built in at every layer and human accountability preserved at every critical decision point.
This shift reinforces a principle that has long guided Visaâs approach to security: resilience must be built in at every layer. Our multilayered security model is designed to detect, prevent, respond, and recover from threats in real time while maintaining the reliability and integrity of our network. As the threat landscape evolves, we are deploying and evaluating advanced defensive AI capabilities to further enhance our security posture and apply those learnings to strengthen protections across the ecosystem we serve.
Attackers are already using these tools. The organizations that move early to deploy defensive AI with this level of discipline will be the ones that hold the line. For Visa, the goal is simple and constant: keep the global movement of money secure for everyone who depends on it.
For a deeper look at Visa's learnings from Project Glasswing, and what this means for cyber resilience, explore the full technical briefing below.