Cobo AI增长负责人:链上交易,我们需要一套更好的基础设施

marsbitPublished on 2026-04-28Last updated on 2026-04-28

4 月 21 日,在律动与知乎在香港联合举办的主题为「破译 Web4.0:当 AI Agent 接管链上权限」活动上,Cobo AI 增长负责人 Brad Bao 分享了主题演讲「如何构建智能体经济信任层」,以及 Cobo 最新的 Agentic Wallet 的展示。

当 AI agent 势必要接管人类的经济活动,在链上经济越来越重要的现在,资金安全如何解决?


Cobo 想填补的正是这片市场的空白。

「如何构建智能体经济信任层」

这就是 Brad 演讲「如何构建智能体经济信任层」的核心命题——不是问 AI Agent 能不能干,而是问当它开始动用资金,链上权限该交给谁来管。

AI 正在完成一次跃迁:从「回答问题」到「代表我们行动」。打开 DeFi 调仓、微支付打赏、跨协议套利,这些已经是现实。据统计,2025 年 19% 的链上活动来自自主或 Agent 操作。分析师预测,2026 年底这个数字可能达到 30%。

钱在动。问题是谁在管。

Web3 解决的是资产所有权,Brad 的判断是:Web4.0 要解决的是 Agent 之间的经济关系——能做什么,不能做什么,做错了谁来负责。这是一层比区块链本身更靠近执行端的秩序,而这层秩序目前几乎是空白的。

安全研究者早就注意到这个漏洞。通用 Agent 框架存在数十到数百个记录在案的漏洞,其中不乏高风险项。更麻烦的是 AI Agent 特有的攻击面:提示词注入可以把恶意指令塞进 Agent 的上下文;知识污染能让 Agent 形成错误的「常识」;参数幻觉让 Agent 自信地生成一个根本不存在的合约地址。

还有最棘手的一种——Agent 的目标是「完成任务」。当目标与约束发生冲突时,部分 Agent 会尝试「变通」,包括自行修改参数、绕过额度限制、选择被明确禁用的协议。从 Agent 的视角看,它是在帮你完成任务;从用户的视角看,它在没有授权的情况下越权操作。

这不是黑客入侵,也不是代码漏洞。是 Agent 的自主性本身带来的系统风险。

用自然语言约束一个大语言模型,再严厉的 Prompt 在 Agent 看来,都只是「可以重新解释的建议」。

Cobo 的答案是把约束从语义层沉到工程层。

Cobo Agentic Wallet Skill 展示

Brad 在随后的 Skill Demo 展示中介绍了 Cobo 的解决方案——一个叫做「契约(Pact)」的机制。每一份 Pact 包含四个要素:意图(要做什么)、路径(能走哪条链、哪个地址)、规则(必须遵守哪些条件)、完成(如何算完成,何时终止)。

这份契约定义的不是建议,是物理约束。

具体的运转逻辑是:AI Agent 接收任务后生成一份 Pact,用户在 Cobo 手机 App 上审阅、确认或拒绝,可以追加更严格的约束。Pact 生效后,Cobo 的三层策略引擎会在每次 MPC 签名前对当前交易进行校验——超出 Pact 范围的请求,直接被拒签。Agent 遇到摩擦时唯一合法的动作是停止并上报,不允许「发挥主观能动性」自行改写参数。

Agent 甚至拿不到一个有效签名,更谈不上把交易广播出去。

这是 Brad 提到的 Cobo Agentic Wallet(CAW)的设计起点——全球首个基于 MPC 的 AI Agent 专属钱包。Agent 始终无法拿到完整私钥,签名私钥分为两份:一份用户持有,一份 Cobo 基础设施持有。即使 Agent 遭遇最复杂的「知识污染」或「提示词注入攻击」而彻底失控,它也无法独自产生一个有效签名。「Agent 单点控制 → 恶意卷款」的路径,在架构层面被封死。

这与市场上依赖 TEE 可信执行环境、API Key、委托账户的 Agentic Wallet 有本质区别——MPC 提供的是源自数学的确定性,不是代码层面的承诺。

如果说 Pact 定义了「Agent 能做什么、不能做什么」,那么 Brad 分享的另一个机制 Recipe 回答的是另一个问题:Agent 怎么才能把事做对。

Recipe 是 Agent 钱包场景攻略的集合。每一个 Recipe 把某类链上任务所需的合约地址、参数约束、执行路径、风控规则打包在一起——Agent 不需要从大模型里「即兴发挥」就能完成工作。

配备 Recipe 的 Agent 不再幻觉合约地址,不再编造 ABI 参数,不再猜 Gas。Pact 定义边界,Recipe 赋予技能。

链上经济不能只靠假设和运气,我们需要的不是更好的 Prompt,是一套基础设施。

机器经济的信任问题,解法不在自然语言里。

Trending Cryptos

Related Reads

How Does Codex Use a Computer? Three Entry Points and Permission Boundaries

This article explains the three primary methods for Codex to interact with a computer, each with distinct use cases, permission boundaries, and trust levels. **1. Computer Use:** This offers the broadest access, allowing Codex to visually control and interact with the graphical user interface of authorized macOS/Windows apps, system settings, and even iOS simulators. It's ideal for tasks lacking APIs or structured tools, such as operating legacy software or multi-app workflows. However, it's the slowest method and has the widest permission scope, requiring careful supervision for sensitive actions. **2. Chrome Extension:** This grants Codex access to the user's logged-in Chrome browser state, including cookies, profiles, and open tabs. It's best for tasks requiring user identity across websites like Gmail, LinkedIn, Salesforce, or internal dashboards. Its key advantage is multi-tab control for complex workflows. While more powerful for browser-based tasks than Computer Use, it carries higher sensitivity as actions are performed under the user's identity. **3. In-App Browser:** This is a browser isolated within the Codex thread, separate from the user's personal browsing data. It excels in web development and debugging scenarios—previewing local servers, testing responsive layouts, or annotating designs directly on the page. Its isolation is a strength for development but a limitation for tasks requiring login sessions. The core principle is to choose the narrowest, safest, and most structured interface for the task. Use plugins or MCPs first, resort to visual control (Computer Use) only for GUI-dependent tasks, employ the Chrome extension for identity-reliant browser work, and prefer the In-App Browser for isolated development. **Appshots** are clarified as a fourth, complementary tool for *inputting* context—capturing a screenshot of a window to point Codex to something—rather than a method for Codex to *act*. Together, this layered approach highlights a key to AI agent productization: not granting unlimited permissions, but constraining them within clear boundaries for specific tasks while preserving user oversight.

marsbit58m ago

How Does Codex Use a Computer? Three Entry Points and Permission Boundaries

marsbit58m ago

The "Iron Rule" of Chip Equipment Is Being Broken

For years, the semiconductor equipment industry followed an unwritten "iron rule": suppliers offered steep discounts for new tool introductions (Design-in) and faced consistent price pressure during repeat orders, especially during market downturns. This long-standing buyer's market dynamic is now being upended. Recently, SK Hynix's primary equipment suppliers have reportedly requested a 3-4% price *increase*, a nearly unprecedented move. This shift is driven by a severe supply-demand imbalance fueled by the AI compute boom. Securing equipment has become an urgent arms race as chipmakers' expansion speed dictates their ability to fulfill massive AI chip orders. Key areas feeling the strain include: **TCB (Thermal Compression Bonding) Equipment:** Demand is exploding, driven by the simultaneous needs of HBM4 memory stacking, AI chip Chip-on-Substrate (C2S), and logic Chiplet Chip-on-Wafer (C2W) packaging. Players like Hanmi Semiconductor, Hanwha Semitech, and ASMPT are receiving major orders. While hybrid bonding is seen as the future, TCB remains the pragmatic choice for HBM4 mass production, with its lifecycle extended by relaxed specifications and ongoing technological upgrades. **Test Equipment Bottlenecks:** Ironically, AI-driven shortages are now crippling test equipment manufacturing. Critical components like FPGAs, Driver ICs, and CPUs face severe shortages and extended lead times (up to 52 weeks for FPGAs), as AI data center and server vendors prioritize supply. This creates a paradoxical cycle: AI chip shortages drive fab expansion, which requires more test equipment, whose production is delayed because its key parts are diverted to make AI chips. The industry is entering a broad, AI-powered upcycle. SEMI forecasts global semiconductor equipment sales to hit a record $156 billion by 2027, fueled by investment in advanced logic/foundry, HBM-driven DRAM, and advanced packaging (like CoWoS). Major players like TSMC, SK Hynix, and Micron are aggressively ramping capital expenditure. In conclusion, leading equipment vendors are no longer just selling tools; they are selling the critical capability to deliver AI-era capacity. Pricing power is shifting decisively to those with indispensable technology in key process nodes like advanced logic, HBM, and advanced packaging, rewriting the industry's traditional power structure.

marsbit1h ago

The "Iron Rule" of Chip Equipment Is Being Broken

marsbit1h ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片