跳到正文
北京时间
原文
PromptArmor:Threat Intelligence·· 2026-07-10精选AI 评分61

PromptArmor 分析:约五分之二 Claude Connectors 会调用外部 AI 服务

2 in 5 Claude Connectors May Call External AI

AI 导读

PromptArmor 审查了当时可用的 487 个 Claude Connectors 共 7517 个工具,发现其中 189 个(约五分之二)的 386 个工具很可能在下游调用其他 AI 服务。

推荐理由

原文给出了 Claude Connectors 二次调用外部 AI 的实测比例、工具分类和具体风险,读者可据此评估企业数据流合规。

正文

Claude

Connectors

Subprocessors

External AI models

Understand which tools call external AI

What does it mean that connectors are calling other AI?

Connectors let Claude take actions in third-party services. As an example, a connector may allow Claude to list emails from Outlook or send an email. However, as more and more apps integrate AI, we've found that third parties developing connectors are now exposing the AI features of their apps as connector tools. The issue is that these AI features may use third-party AI models and subprocessors that may not have been approved by your organization.

We reviewed all 7517 tools from the 487 Claude connectors available at the time of this publication, and assessed the risk of each tool routing requests through a second AI service downstream. Of the 487 connectors, 189, roughly two in five, contained tools that are highly likely to call additional AI services.

What the flagged tools do

The 386 flagged tools split across generation, retrieval, summarization, and domain-specific models.

Content / image / media generation109

Semantic search / RAG retrieval85

Summarization / Q&A81

Recommendation / ranking37

Extraction / document understanding24

Domain-specific ML model10

Other15

Generation is the largest group: more than a quarter of the flagged tools create new content from your input. The rest search, summarize, transcribe, or extract it, and each one routes your data through a model to do it.

What are the risks

Connectors are calling AI tools that you may not have validated when you reviewed the vendor being connected. Each downstream call inherits that AI service's data practices, not just the vendor's.

  • Data residency: the downstream model can run in a region your DPA never approved, and may move without notice.

  • Training data and retention: your inputs may be retained, subject to human review, or used to train AI models.

  • Nth-party data processing: your data is sent from Claude to the connected app and then to an AI subprocessor such as Azure or even a model provider directly, such as the OpenAI API. Each of these providers may have differing data processing agreements.

  • Limited model governance: the model behind a connector's tool can be swapped with no visibility to the connector, impairing an organization's ability to vet whether tools are calling models that have not been approved by one's organization (such as DeepSeek).

As an example, if your Claude agent activates Zoom's connector tool to search meetings with natural language, and passes in a query containing sensitive data, Zoom AI may send that data to any of its ten AI subprocessors in order to generate a response from one of eight different model families it uses.

AI-backed Connector Examples

Here are some examples of mainstream connectors flagged for having tools that leverage AI systems downstream.

Box

/20 tools

Box AI reads your documents to answer questions.

Tools:

AI QA Single File, AI QA Multi File, +3

Glean

/8 tools

Semantic ranking over everything you connected to Glean.

Tools:

Search, Code Search, +4

Adobe for creativity

/67 tools

Generate and segment image content.

Tools:

Generative Expand, Remove Background, +6

Zoom

/5 tools

Natural-language search and Q&A over your meeting recordings and transcripts.

Tools:

Search Meetings

Fun fact

In 54 connectors, at least half of their tools were flagged as AI-backed.

Know which connectors call other AI

来源:PromptArmor:Threat Intelligence · promptarmor.com