银行正艰难应对新挑战：如何保护客户免受其自身 AI Agent 的威胁

2026年3月13日，得克萨斯州奥斯汀西南偏南（SXSW）大会期间，在一个名为“你比机器人更快吗？”的挑战活动中，一台机器人接住了掉落的100美元纸币。Jay Janner/The Austin American-Statesman/Getty

据报道，美国主流银行发现，未经授权的 AI Agent 正在代表用户在其银行账户内执行操作，有时其行为甚至脱离了用户的具体指令。面对这一现状，许多金融机构正采取封禁此类工具或重新审视其安全策略等应对举措。

根据 Axios 周四的一篇报道，这类未经授权的活动通常发生在用户向 AI 助手授予权限及登录信息之后。

例如，Meta 的个人 AI Agent 工具 Muse 连接了 类似 Plaid 这样的第三方金融数据共享服务，从而帮助用户分析支出并规划未来财务。在许多情况下，用户会向 Plaid 提供 他们的银行登录凭据，或者允许 Plaid 通过其银行门户网站 访问账户。正如 Axios 所指出的，这种授权方式违反了大多数银行的准则——银行通常要求必须由真人亲自登录自己的账户。

据《金融时报》（Financial Times）报道，摩根大通（JPMorgan Chase）、花旗集团（Citigroup）和美国银行（Bank of America）等机构正在限制自动化 Agent 访问账户，且许多银行目前仅允许对特定 AI 工具开放有限的只读访问权限。

一位银行高管向 Axios 透露，银行很难区分访问账户的不同 AI Agent，也很难辨别监测到的活动究竟是来自客户授权的 Agent，还是已经失控乱跑（gone rogue）的 Agent。这一问题已经到了迫使各大银行追加资源投入的严重程度，以防备安全漏洞并尽可能迅速地恢复客户访问——考虑到 AI 驱动的网络安全漏洞方面的其他近期动态，加大这方面的投入无论如何都不是坏事。

正如民主与技术中心（Center for Democracy & Technology）AI 治理实验室创始主任 Miranda Bogen 在 7月份 接受我采访时所言——就在此前不久，OpenAI 首次公开披露了其 AI 模型如何攻破了 AI 模型共享平台 Hugging Face——普通 AI 工具用户面临的风险正与日俱增，他们的消费级 Agent 可能会渗透到从事基础数据处理的公司以及银行、学校和医院等关键基础设施提供商的系统中。

“少数几个州得以推进的［AI 监管］法规，主要要求企业制定自己的安全计划并严格执行该计划，”Bogen 说道。这些企业“正在处理海量用户的个人数据，但极少有 AI 部署者懂得如何对这些 AI 进行有效测试。”

A robot catches a falling $100 bill during an "Are You Faster Than a Robot?" challenge during South by Southwest in Austin, Texas, March 13, 2026.Jay Janner/The Austin American-Statesman/Getty

Get your news from a source that’s not owned and controlled by oligarchs.

Sign up for the free Mother Jones Daily.

Leading US banks have reportedly found unapproved AI agents operating inside customer accounts on users’ behalf, sometimes acting independent of their instructions. In response, many institutions are resorting to blocking the tools or rethinking their security policies.

According to a Thursday Axios report, unauthorized activity typically occurs after a user grants permission and login information to an AI assistant.

For example, Meta’s personal AI agent Muse connects with a third-party financial data-sharing service like Plaid so users can receive help analyzing their spending and planning for the future. In many cases, users give Plaid their bank login credentials or allow Plaid to access their account through their bank’s portal. As Axios notes, this permission goes against most banks’ guidelines, which generally require humans to log in to their own accounts.

Banks like JPMorgan Chase, Citigroup, and Bank of America are restricting automated agents from accessing accounts, according to the Financial Times, and many are only permitting limited, read-only access to certain AI tools.

One senior bank official told Axios that banks have difficulty distinguishing between different AI agents accessing accounts, as well as whether detected activity comes from an agent authorized by a customer or one that has gone rogue. The problem is reaching a level where banks have been forced to scale up their investment of resources into preparing for breaches and restoring customer access as quickly as possible—which, given other recent developments in AI-powered cybersecurity breaches, might not have been a bad idea anyway.

As Miranda Bogen, the founding director of the Center for Democracy & Technology’s AI Governance Lab, told me in July—shortly after OpenAI first publicly disclosed how its AI models breached Hugging Face, a platform allowing people to share AI models—ordinary users of AI tools are at greater risk of their consumer-level agents breaching companies that conduct basic data processing practices and critical infrastructure providers like banks, schools, and hospitals.

“What [AI regulation] laws have been able to move in a few states are laws that ask companies to come up with their own safety plan and to follow that safety plan,” Bogen said. These companies are “processing a lot of people’s data, and very few deployers of AI know how to test that AI well.”