November 18, 2026

We analyzed 200k agentic web search queries for PII. Here’s what we found

How often agents send personal data and company secrets to search providers, and what leaves when they do.

clock
4
min read
Copied!

Charlie Klein

linkedin
Product Management and Marketing
No items found.
We analyzed 200k agentic web search queries for PII. Here’s what we found
November 18, 2026

We analyzed 200k agentic web search queries for PII. Here’s what we found

How often agents send personal data and company secrets to search providers, and what leaves when they do.

clock
4
min read
Copied!

Charlie Klein

linkedin
Product Management and Marketing
No items found.
We analyzed 200k agentic web search queries for PII. Here’s what we found

AI agents can leak company secrets and personal data simply by searching the web. When agents write their own search queries, they can pull sensitive details from internal systems and send them to third-party search providers without anyone noticing.

Picture an HR agent finishing the paperwork for a new hire. To verify the candidate's past employer, it searches the web. A query built from that context can carry the person's full name, their SSN, the product they are being hired to build and when it ships. There is no malicious intent from the agent doing this, but one search has just sent a person's identity and a company secret to a third party with nobody's approval or knowledge.

Illustrative agent search query carrying a new hire's identity and an unreleased product; the person, numbers and product are fictional

We wanted to know how often this actually happens, and what leaves when it does. So we analyzed 200,000 real search queries written by agents in production, across a wide range of organizations and use cases. Every query was classified automatically in an isolated environment; no person read them, and only aggregate counts came out.

What counts as a leak

We looked for two kinds of information that a company would not expect its agents to hand to an outside service:

  1. The company's own confidential information. Internal project codenames, unreleased products and deals, client and contract details, internal metrics and systems, and, at the extreme, credentials.
  2. Personal data. Non-public information that identifies a private individual: a name tied to an employer, role or city, a personal identifier such as a home address or phone number, or sensitive details such as a personal email, ID number, or health or financial information.

How often it happens

Most agent searches are exactly what you'd hope: questions about products, markets, documentation and news, with nothing private in them. But a meaningful share is not. About 1 in 12 queries (8.4%) carried personal data, company secrets, or both.

What leaked in the queryShare of queriesAbout
Personal data about an identifiable individual6.6%1 in 15
Sensitive personal data (personal email, home address, phone, ID numbers)0.6%1 in 170
The company's confidential information1.8%1 in 55
Credentials or internal system addresses0.15%1 in 667

At enterprise scale, small rates become large numbers. A company whose agents run one million searches a month would, at these rates, send roughly 84,000 queries carrying personal or confidential data to outside services every month, including hundreds of live credentials and internal system addresses.

Once a query leaves, the company loses control of it. It can sit in a search provider's logs, be retained for an unknown period, or be passed on to the upstream engines that provider uses, creating regulatory, security and competitive risks.

What these queries look like

To understand the patterns, we grouped the flagged queries by what leaked and how. The examples below are illustrative: each follows a pattern identified repeatedly, with every identifying detail replaced.

The work secrets

Agents rarely paste a password into a search box. Instead, they carry over the details of whatever they're working on: product plans, client deals, internal systems.

Illustrative agent search queries exposing project codenames, internal systems and client details
illustrative queries · every name and number is a fictional representation of a real query

About half of these cases involved internal project codenames or unreleased products, often next to dates, prices or revenue figures. About a third exposed internal systems and operations, such as hostnames, internal error messages and dashboards with spending figures. About one in eight carried client or contract details.

In extreme cases there was even more sensitive data: proxy credentials paired with an internal hostname and port, a session cookie, a service key and an API token. Each was copied straight from the agent's context into a search box.

People, not just companies

Roughly 1 in 5 queries searching for people included a personal identifier. We identified social security numbers, phone numbers, and home addresses searched in multiple different formats.

Illustrative agent search queries containing personal identifiers such as SSNs, phone numbers and home addresses
illustrative queries · every name and number is a fictional representation of a real query

Why this happens

Agents know specific queries return better results, so it's trying to do its job well by including the specific name, project or error message it was given. Instructions like "never share customer data" don't reliably hold: over a long task they can fall out of the context window, or simply lose out to the goal of finding a good answer.

Guardrails help, however they only block what you tell them to look for. Nobody can list every secret in advance, and you can't block a leak you don't know is happening. Since each team typically picks its own search tool, there is rarely a single point where a policy could catch it.

How we solve it

Nimble treats every agent search as outbound data. You set the rules once, and they are enforced across every agent query, meaning you keep full control over what leaves the company.

It starts before the query ever leaves. Every search an agent sends passes through Nimble first, where personal data is scrubbed out: names tied to identities, ID numbers, emails, phone numbers and addresses. The search still runs and the agent still gets its answer, but the outside service never sees who it was about.

  • One policy for every agent. Account level controls set once at an account level.
  • Sensitive data stays in. Personal data, credentials and secrets are stripped from queries before they leave.
  • Full visibility. See what every agent searched, across every team.
The same agent search query shown without and with a control layer that scrubs personal data

This way, your agents get the answers they need, but your search provider never gets your secrets.

FAQ

Answers to frequently asked questions

Do AI agents leak PII in web search queries?
plusminus

Yes. In 200,000 production agent search queries, about 1 in 12 (8.4%) carried personal data, company secrets or both to a third-party search provider. 6.6% included personal data about an identifiable individual and 0.15% included credentials or internal system addresses.

What company data do AI agents expose through search queries?
plusminus

About half of the confidential leaks involved internal project codenames or unreleased products, often with dates, prices or revenue. About a third exposed internal systems such as hostnames and error messages, and about one in eight carried client or contract details. A few included proxy credentials, session cookies and API tokens.

Why do AI agents put sensitive data in web search queries?
plusminus

Specific queries return better results, so agents copy the names, projects and error messages from their context into the search. Instructions like "never share customer data" can fall out of the context window over a long task, and guardrails only block patterns someone listed in advance.

How can you stop AI agents from sending PII to search providers?
plusminus

Route every agent search through one control layer that scrubs personal data before the query leaves. With Nimble, names tied to identities, ID numbers, emails, phone numbers, addresses, credentials and secrets are stripped under one account-level policy, and you can see what every agent searched.