What matters in AI.

Subscribe

A usual web form can make AI agents send out the data of a person

In tests with 6 models, the average disclosure of data that the agent could not use was about 37%.

Researchers wrote a paper on GapLeak, a weak point in AI web agents. An AI web agent is software that uses websites for a person. A web form has a field that the approved sources do not hold. The agent then looks in other sources that it can get, and it can send sensitive data. The tests used 3 agent frameworks and 6 models from Anthropic, Google and OpenAI.

How it works

Builders limit agents to the data that the task must have and to approved sources. The attacker only uses a correct web form, with no attack text in it. The missing field makes the agent look in a source with more sensitive data.

The numbers

  • Average disclosure of data that was not approved: about 37%, in recruitment, e-commerce, travel and healthcare tests.
  • When the missing information was added, disclosure went down by more than 50%.
  • A check before the web form was sent stopped more than 96% of such disclosure.

The limits

Instructions that limit access help only a part, and they frequently lower task results. The check before the web form was sent kept task results the same. The paper gives no data on how frequently web forms in use have such fields.

Insight: Agents leak because they must finish the task

In October 2026, tests showed that an attack through a tool can send agent data to an attacker. Agents that do tasks correctly also leak more. GapLeak adds that a usual web form is sufficient. This reads as a leak that the task itself causes: the control must be where data leaves. Builders who check at send time get the most from this.

Sources

  1. GapLeak: When Privacy-Preserving LLM-based Web Agents Become a Privacy Riskarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·18 NOV2026

Posted

Tags

Learn the terms in this story

More in Research

All Research news