We have a year to fix security everywhere
309 points
• 6 days ago
• Article
Link
最近发布的 GLM 5.3-flash 是一款开源权重的大型语言模型，能在消费级硬件上快速且低成本运行，这打开了一个关键的脆弱性窗口。因为该模型可以被修改以移除安全防护，任何在硬件上有少量投入的人现在都能部署一个强大且无约束的工具，用来实施复杂的网络攻击。这使风险从理论上的可能性变成了迫在眉睫的现实威胁，恶意活动可以被大规模自动化。
危险在于模型的可及性与其在网络安全任务上出色性能的叠加。基准测试显示，GLM 5.3-flash 在识别并利用现实世界漏洞方面表现优异，几乎把先进的自动化黑客能力交到了任何人手中。尽管 Project Glasswing 和 Daybreak 等项目正尝试用前沿模型修补安全缺陷，但核心难题仍是如何在分散且常为遗留系统的基础设施中迅速部署这些补丁。
事态需要紧急应对，因为攻击者的推进速度已快于防御方。政府和监管机构应将重心从仅仅发现漏洞转向强制并激励补丁的实际部署。政策应优先支持基于风险的定期渗透测试，并对关键基础设施——例如电网、金融系统和电信——实施更严格的安全标准。单靠对模型或硬件的禁令很可能无效且适得其反，因为技术无可避免地会扩散。
企业和开源基金会应把这视为利用前沿 AI 强化防御、全面改造安全态势的良机。这需要超越手工安全流程，采用结构化提示、自动化测试和形式化验证来加速修补周期。应赋能内部安全工程团队，优先进行漏洞的筛选与补丁部署，确保在漏洞被自动化代理利用前得到修复。
最后，组织需建立健全内部控制，例如沙箱隔离、限定权限的凭证和纵深防御，以防自有防御代理成为事故源。逐步放弃内存不安全的语言、全面清点依赖项并演练快速的事件恢复，都是必要步骤。即便对当前威胁持怀疑态度的人，也应看到这些措施代表了网络安全必须尽快实现的进步，应在机会窗口关闭前付诸行动。
The recent release of GLM 5.3-flash, an open-weight large language model that is both fast and cheap to run on consumer hardware, has created a critical window of vulnerability. Because the model can be modified to remove safety safeguards, anyone with a modest investment in hardware can now deploy a powerful, unconstrained tool capable of performing sophisticated cyberattacks. This capability moves the needle from theoretical risk to an imminent, real-world threat where malicious activities can be automated at scale.
The danger lies in the combination of the model's accessibility and its high performance in cybersecurity tasks. Benchmarks indicate that GLM 5.3-flash is highly proficient at identifying and exploiting real-world vulnerabilities, effectively putting the power of advanced, automated hacking into the hands of virtually anyone. While existing initiatives like Project Glasswing and Daybreak have been working to patch security flaws using frontier models, the core challenge remains the slow and difficult process of deploying those fixes across fragmented, often legacy, infrastructure.
Urgency is required to address this landscape, as attackers are currently advancing faster than defenders. Governments and regulatory bodies should shift their focus toward mandating and incentivizing the actual deployment of patches, rather than just the identification of vulnerabilities. Policy should prioritize frequent, risk-based penetration testing and enforce stricter security standards for critical infrastructure, such as power grids, financial systems, and telecommunications. Relying on bans for models or hardware is likely to be ineffective and counterproductive, as the technology will inevitably proliferate regardless of such measures.
Companies and open source foundations must treat this as a prime opportunity to overhaul their security posture by leveraging frontier AI as a defensive tool. This involves moving beyond manual security practices and utilizing structured prompts, automated testing, and formal verification to accelerate the patching cycle. Internal security engineering teams should be empowered to prioritize the triage and deployment of fixes, ensuring that vulnerabilities are addressed before they can be exploited by automated agents.
Finally, organizations need to implement robust internal controls, such as sandboxing, scoped credentials, and defense-in-depth, to prevent their own defensive agents from becoming vectors for incident creation. Moving away from memory-unsafe languages, maintaining thorough inventories of dependencies, and practicing rapid incident recovery are essential steps. Even for those skeptical of the immediate threat level, these measures represent a long-overdue advancement in collective cybersecurity that must be acted upon before the window of opportunity closes.
367 comments • Comments Link
• 关于 Large Language Models (LLMs) 在提供诸如制造管式炸弹（pipe bombs）等危险行为指南方面所带来的威胁存在广泛争议。一方面有人认为这类信息历来容易获取；另一方面有人指出，降低获取门槛会显著增加冲动个体实施危险行为的风险。
• 虽然物理威胁更难追溯，但 AI 的能力正在加速网络和生物领域的威胁演进。有人认为 AI 对攻防双方都有利，但也有人强调一种内在的不平衡：攻击者只需一次成功，而防御者必须持续、近乎完美地防守。
• 对"Moore's Law of Mad Science"持怀疑态度的人指出，该理论低估了巨大的后勤和技术障碍，比如获取受限原料或高端实验设备；即便有超强智能辅助，普通人也难以轻易绕过这些障碍。
• 部分观点认为 AI 是一种催化剂，会从根本上改变而非彻底摧毁安全态势——正如以往技术促使人类适应一样，社会也会相应调整防御策略。
• 一个主要争论是，社会是否正处于所谓的"Finding Out"阶段：过去忽视安全最佳实践会不会导致系统性崩溃，还是这些担忧大多被过度渲染。
• 一些硬件进展（例如 memory tagging 和计算密集型的 local AI hardware）被视为未来必要的护栏，但它们被采用得很慢，这令安全从业者颇为沮丧。
• 对庞大且未经充分审计的 dependency graphs，以及对像 WordPress 这样脆弱的软件生态的依赖，被视为关键弱点。极简主义者主张回归 static site generation 、尽量减少第三方代码，这被看作迈向真正安全的重要第一步。
• 行业内普遍存在一种沮丧感：在当前以速度、便利和功能为先的商业激励下，安全常被当作一种形式化的"打钩"流程，而非必须达成的技术目标。
• 关于 memory-safe languages 和更安全的 OS architectures（例如 microkernels）是否必要的争论依然存在，许多人认为在依赖大量 legacy 部署的现实中，这些方案短期内并不现实。
• 这场辩论凸显了两种根本不同的看法：一方认为安全是可以通过技术手段控制的挑战；另一方则认为软件行业的组织与社会结构本质上难以优先投入那种必要的、长期的安全建设。
总体来看，这场讨论反映了 AI 驱动的安全加速潜力与现代软件基础设施内在脆弱性之间的张力。普遍共识是，AI 为发现和利用漏洞带来了强大新能力，但根本性的问题（如 technical debt 、缺乏严格审计和错位的商业激励）已存在数十年。至于这会在不久将来引发灾难性的清算，还是引导出适应与加固的常规周期，尚无定论；但可以肯定的是，在 AI 驱动的自动化漏洞研究时代，依赖"box-checking"式的安全和臃肿、不透明的软件堆栈正变得越来越不可持续。 • The threat posed by Large Language Models (LLMs) in providing instructions for dangerous acts like building pipe bombs is often debated, with some arguing that such information has long been easily available while others emphasize that lowering the barrier to entry significantly increases the danger from impulsive individuals.
• While physical threats may be harder to source, AI capabilities accelerate cyber and biological threats. Some suggest that AI benefits defenders as much as attackers, but others note an inherent imbalance where attackers only need one success to win, whereas defenders must be perfect indefinitely.
• Skepticism exists regarding the "Moore's Law of Mad Science," which suggests the intelligence required to cause global harm is rapidly dropping. Critics argue this ignores the massive logistical and technical hurdles—such as accessing restricted materials or high-end laboratory equipment—that even super-intelligent assistance cannot easily circumvent for the average person.
• AI is perceived by some as a catalyst that will fundamentally change rather than destroy security, suggesting that just as previous technologies forced adaptation, society will evolve its defense strategies.
• A significant point of contention is whether society is currently in a "Finding Out" phase where past negligence regarding security best practices will lead to systemic failures, or if the fears are largely overblown hype.
• Hardware advancements, such as memory tagging and compute-heavy local AI hardware, are viewed as essential future guardrails, though their slow adoption rates remain a point of frustration for security-minded practitioners.
• The reliance on sprawling, unaudited dependency graphs and fragile software ecosystems like WordPress is identified as a critical vulnerability. Proponents of simplicity suggest that returning to static site generation and minimizing third-party code is a vital first step toward true security.
• A prevailing frustration in the industry is that security is often treated as a bureaucratic checkbox rather than a technical imperative, primarily because current business incentives prioritize speed, convenience, and features over robust protection.
• Arguments for the necessity of memory-safe languages and more secure OS architectures like microkernels persist, though many view these as impractical solutions for the current reality of widespread, legacy-dependent software deployment.
• The debate highlights a deep divide between those who believe security is a manageable technical challenge and those who believe the current organizational and social structures of the software industry are structurally incapable of prioritizing the necessary, long-term security investments.
The discussion reflects a broad tension between the potential for AI-driven security acceleration and the inherent fragility of modern software infrastructure. A prevailing sentiment is that while AI introduces powerful new capabilities for both finding and exploiting vulnerabilities, the fundamental security problems—such as technical debt, lack of rigorous auditing, and misaligned business incentives—have existed for decades. Consensus remains elusive regarding whether the near future will bring a catastrophic reckoning or a standard cycle of adaptation and hardening, though there is a clear acknowledgment that current reliance on "box-checking" security and bloated, opaque software stacks is increasingly unsustainable in an era of automated, LLM-powered vulnerability research.