AI智能体自主行动时出问题的5件事

当 AI 智能体开始执行真实世界中的操作时,会出现五大治理问题,涵盖访问权限失控、身份泛滥,到审批瓶颈与人工监督。
作者:KDnuggets,2026年11月16日,发表于 Partners 栏目
一个智能体被要求调查某个数据质量问题。它追踪了数据管道并找到了根因。随后,它把 1800 条敏感客户记录导出到一个共享工作区,“以便团队核实”。没有人要求它这样做,而按它自己的衡量标准,它成功了。
这个场景只是举例说明,但如今智能体确实会写入生产系统。随之而来的是五个问题。
1. 模型已经足够好,落地却卡住了。
在 OSWorld 基准测试上的任务完成率从 2024 年的约 12% 提升到 2026 年的 66%。然而,在任何一个单一职能中规模化部署智能体的企业都不超过 10%。大多数团队说不清自己的智能体被允许接触什么。
2. 大部分损害都始于内部。
80% 的组织报告过智能体的风险行为,例如不当数据暴露。这些案例中很少涉及攻击者。一个销售团队克隆了报价智能体并修改了提示词,这个副本就一直在套用法务部门一个季度前已废止的折扣规则。
3. 你的智能体清单是错的。
每名人类对应约 45 个非人类身份。不到一半的 CISO 知道自己的智能体在哪里、能访问什么。1 一名开发者用一个周末把智能体接入 CRM,然后在 3 月离职。这个智能体仍以他的凭据继续运行。每个智能体都需要自己的身份,以及一位仍在职的负责人。
4. 逐个智能体审批是死路一条。
如果你的流程最终是由委员会为每个新智能体签字放行,那就放弃它。在 10 个智能体时这套还行得通。到了 100 个就会崩溃,更何况如今智能体已随你付费购买的软件一同交付。机器可读的智能体契约更经得起考验:每个智能体都继承一套基线规则,再按风险和领域叠加更细的规则。
5. 人工审查已经在吞噬你的团队。
Harris Poll 调查了 306 位美国数据、隐私和 AI 决策者。87% 表示,他们的团队要花数小时反复核验智能体所依赖的上下文。Guardian 智能体会在每个动作运行时比对其契约进行核查,然后标记、重定向或阻止该动作。人只会看到越过红线的那部分。其余一切照常运行。
下载这份 playbook,获取 Scan、Set、Supervise 框架和一份 10 步行动计划。
参加 AI Trust Score 评估,看看应当先解决哪个问题。
申请 Collibra AI Command Center 的演示。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(5 Things That Break When AI Agents Start Acting on Their Own)
Five governance problems emerge when AI agents begin taking real-world actions, from unmanaged access and identity sprawl to approval bottlenecks and manual oversight.
By KDnuggets on November 16, 2026 in Partners
An agent is asked to investigate a data quality issue. It traces the pipeline and finds the root cause. Then it exports 1,800 sensitive customer records to a shared workspace “so the team can verify.” Nobody asked it to, and by its own measure it succeeded.
That scenario is illustrative, but agents do write to production systems now. Five problems follow.
1. The Models Are Good Enough. Rollouts Are Stuck.
Task completion on the OSWorld benchmark went from about 12% in 2024 to 66% in 2026. Yet no more than 10% of enterprises are scaling agents in any single function. Most teams can’t say what their agents are allowed to touch.
2. Most of the Damage Starts Inside.
80% of organizations report risky agent behavior such as improper data exposure. Few of those cases involve an attacker. A sales team clones a quoting agent and edits the prompt. The copy keeps applying discount rules legal retired a quarter ago.
3. Your Agent Inventory Is Wrong.
There are about 45 nonhuman identities for every human. Fewer than half of CISOs know where their agents are or what they can access.1 A developer wires an agent into the CRM over a weekend and quits in March. The agent keeps running on their credentials. Every agent needs its own identity and an owner who still works there.
4. Per-Agent Approval Is a Dead End.
If your process ends with a committee signing off on each new agent, drop it. That works at ten agents. It collapses at a hundred, now that agents ship inside software you already pay for. Machine-readable agent contracts hold up better: one baseline every agent inherits, plus narrower rules by risk and domain.
5. Manual Review Is Already Eating Your Team.
A Harris Poll surveyed 306 US data, privacy and AI decision makers. 87% said their teams spend hours re-verifying the context their agents rely on. Guardian agents check each action against its contract while it runs, then flag, redirect, or block it. People see only what crossed a line. Everything else just runs.
Download the playbook for the Scan, Set, Supervise framework and a 10-step action plan.
Take the AI Trust Score assessment to see which problem to fix first.
Request a demo of Collibra AI Command Center.
原文为英文,正文由自动抓取并保留原文。
原文发布于 KDnuggets(5 Things That Break When AI Agents Start Acting on Their Own),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。