密码管理服务商1Password近日推出全新的“1Password for Agents”功能，允许人工智能智能体在获得用户授权后，使用用户保存的账户凭据登录网站并执行任务，同时避免将密码直接暴露给AI模型。这项功能旨在解决AI智能体日益普及后面临的一项重要安全问题：当AI需要代表用户登录网站、操作账户时，如何让它完成任务，又不必将用户的敏感凭据交给它。
随着AI智能体逐渐从回答问题转向执行实际操作，用户开始让它们处理需要登录账户的复杂任务，例如管理在线服务、访问工作平台或完成跨网站操作。但传统方式通常要求用户将密码交给AI，或者在智能体每次需要登录时亲自接管浏览器。这两种方式分别存在凭据暴露风险和操作效率低下的问题。
1Password希望通过新的授权机制改变这一模式。按照该公司的设计，AI智能体可以请求使用特定账户的登录凭据，但不能直接读取密码本身。用户仍然掌握授权权力，可以决定是否允许智能体使用某项凭据，而不是将整个密码库开放给AI。
具体而言，当AI智能体访问某个需要登录的网站时，它会使用唯一的数字密钥向1Password证明自己的身份。随后，1Password会向用户发出通知，说明智能体需要使用哪个账户的凭据，以及请求授权的原因。用户可以在自己的设备上批准或拒绝该请求。
如果用户批准，1Password便会将所需的登录信息直接填入网站的登录界面，而不是将密码以文本形式发送给AI智能体。智能体随后可以继续执行任务，但不会直接看到或处理用户的实际密码。
这种设计与传统的双重身份验证具有一定相似性：敏感操作需要用户主动确认。不过，两者并不完全相同。双重身份验证主要用于确认登录者的身份，而1Password for Agents的核心目标，是让用户能够对AI智能体使用账户凭据的行为进行授权和控制。
1Password首席执行官阿希什·特里帕蒂表示，AI智能体已经开始替用户完成实际工作，过去那种直接交出凭据、再期待系统安全运行的方式已不再适合这一发展趋势。他认为，智能体应该能够完成用户交代的任务，却不应因此获得永久访问用户所有账户的权限。
根据1Password公布的信息，该功能将支持多种AI智能体和工具，包括Meta Muse、ChatGPT Work、ChatGPT Dots、Claude Duet、Grok Bot以及Town等。与此同时，Browserbase、Browser Use、Dana、Instinct、Kernel、Perplexity、Ramp、Tomo和Wajo等平台也正在推进相关集成，未来预计还会有更多合作伙伴加入。
这意味着，1Password希望将自身从传统的密码管理工具，进一步发展为AI智能体的身份验证和访问控制层。当AI不再只是生成文本，而是开始代表用户登录网站、读取信息并执行操作时，如何确认智能体身份、限定其权限，以及追踪其使用凭据的行为，将变得越来越重要。
不过，这种机制并不意味着AI智能体可以在完全没有风险的情况下操作账户。即使智能体无法直接看到密码，它仍然可能在获得授权后执行用户允许的操作。如果AI被恶意网页内容误导，或者受到提示注入攻击影响，就可能尝试进行用户并不希望发生的操作。因此，用户仍需要仔细审查授权请求，并对涉及资金、个人信息或重要账户设置的任务保持谨慎。
1Password此前也曾提醒用户注意AI辅助浏览带来的安全风险。例如，如果浏览器中的AI助手能够读取不可信网页内容，而1Password扩展程序又处于解锁状态，恶意内容就可能诱导智能体尝试执行用户原本能够执行的浏览器操作。对此，1Password建议用户关闭不必要的自动登录功能，并在浏览不可信内容或离开设备时锁定扩展程序。
此次推出的AI智能体凭据授权功能，正是试图在自动化效率与账户安全之间建立更明确的边界。它的核心理念不是让AI获得更多秘密，而是让AI在不接触秘密的情况下，获得完成特定任务所需的有限权限。
随着AI智能体逐渐进入办公、购物、账户管理和企业业务流程，传统密码管理工具也可能需要重新定义自身的作用。未来，密码管理服务不再只是帮助用户保存和自动填写密码，还可能负责管理哪些AI可以访问哪些账户、每次授权的具体范围，以及授权何时失效。
对用户而言，这种模式有望降低直接向AI提供密码所带来的风险，同时减少执行复杂任务时反复手动登录的麻烦。但它能否真正成为AI智能体的通用安全标准，仍取决于更多平台的接入情况、授权机制的可靠性，以及不同AI系统在实际使用中的安全表现。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The password manager, 1 Password, recently introduced a brand-new “1Password for Agents”, allowing artificial intelligence AI agent, after obtaining user authorization, to log on and perform tasks using user-held accounts, while avoiding direct exposure to AI models. This function is designed to address an important security issue in the wake of the growing spread of AIAI agent: how to get AI to do its job when it needs to access the website on behalf of its users and operate the accounts without having to hand over to it the user ' s sensitive documentation.
As AIAI agent gradually moved from answering questions to implementing operations, users began to allow them to handle complex tasks requiring access to accounts, such as managing online services, accessing work platforms or completing cross-site operations. However, traditional methods typically require users to give the password to AI or take over the browser in person whenever AI agent is required to log in. The two approaches are characterized by evidence-based exposure risks and inefficiencies.
1 Password hopes that this model will be changed through a new delegation of authority mechanism. According to the design of the company, AZAI agent could request the use of login documents for specific accounts, but could not directly read the password itself. Users remain empowered to decide whether to allow AI agent to use a certificate rather than to open the entire password library to AI.
Specifically, when AZAI agent visits a website that needs to be accessed, it uses the only digital key to prove its identity to one Password. Subsequently, 1 Password will send a notice to the user indicating which account AI agent is required to use and the reasons for the request for authorization. Users may approve or reject the request on their own equipment.
If approved by the user, 1 Password will enter the required login information directly into the web portal instead of sending the password in text to AIAI agent. AI agent could then continue the mission without seeing or processing the user ' s actual password directly.
This design is somewhat similar to the traditional dual identification: sensitive operations require active user confirmation. However, they are not identical. Dual identification is primarily used to confirm the identity of the login, while the central objective of 1 Password for Delegates is to enable users to authorize and control the use of AZZ AI agent to document the use of accounts.
1 Password Chief Executive Officer, Ahish Trepati, stated that AZAI agent had begun to perform practical work on behalf of the users and that the way in which the system had operated safely, by handing over the documents in the past, was no longer appropriate for this development. In his view, AI agent should be able to perform the client ' s task, but should not be granted permanent access to permission for all user accounts.
According to information published by 1 Password, this feature will support a variety of AIAI agent and tools, including Meta Muse, ChatGPT Work, ChatGPT Dots, Claude Duet, Grok Bot and Town. At the same time, platforms such as Brownserbase, Browner Use, Dana, Instinct, Kernel, Perplexity, Ramp, Tomo and Wajo are advancing the integration and are expected to include more partners in the future.
This means that 1 Password wants to develop itself from the traditional password management tool to the AZAI agent identification and access control layer. It will become increasingly important for AI to identify AI agent, qualify permission and track its use of documents when it no longer produces text, but begins to log in on behalf of users, read information and perform operations.
However, this mechanism does not mean that AZAI agent can operate the accounts without any risk whatsoever. Even if AI agent does not have direct access to the password, it may carry out user-authorized operations after obtaining authorization. If AI is misled by malicious web content or is affected by a reminder of an attack, it may attempt to perform an operation that the user does not wish to do. As a result, users still need to scrutinize requests for authorization and exercise caution with regard to tasks involving funds, personal information or critical account set-up.
1 Password has also previously alerted users to the security risks associated with AI-assisted browsing. For example, if the AI assistant in the browser is able to read untrustworthy web content and the 1Password extension is unlocked, malicious content may induce AI agent to try to perform a browser operation that the user could otherwise perform. In response, 1 Password recommended that users close unnecessary automatic login and lock the extension when browsing untrustworthy content or leaving the device.
This launch of AIAI agent is an attempt to establish a clearer boundary between automation efficiency and account security. Its core idea is not to give AI more secrets, but to give AI the limited permission required to perform a particular task without touching them.
As AIAI agent moves into office, shopping, account management and business processes, traditional password management tools may need to redefine their role. In the future, the password management service will no longer only help users to keep and automatically fill in passwords, but may also be responsible for managing which accounts may be accessed by AI, the specific scope of each authorization and when the authorization will expire.
For users, this model is expected to reduce the risk of providing passwords directly to AI, while reducing the difficulty of repeated manual log-ins in the execution of complex missions. However, whether it truly serves as a common security standard for AIAI agent still depends on access to more platforms, the reliability of the authorization mechanism and the security performance of different AI systems in practice.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.