Claude’s Share Feature Leaks API Keys, Resumes, and SSNs via Google Indexing

Background

Anthropic’s Claude AI assistant offers a “share chat” feature that generates a URL for others to view a conversation. The pages are served without a noindex directive, so search‑engine crawlers can index them.

Data exposed through Google search

Security researchers found that Google can return many shared Claude pages containing highly sensitive information, including:

Credentials: API keys, crypto‑wallet addresses, database passwords.

Personal identity: real names, home addresses, phone numbers appearing in résumés.

Professional sensitive data: lawyers’ case notes, engineers’ internal project details.

High‑sensitivity data: suspected Social Security Numbers (SSNs).

Private conversations: personal chat logs not intended for public view.

Why the issue persisted

The same omission occurred a year earlier with ChatGPT; OpenAI patched it quickly after reports, but Anthropic left the flaw unaddressed for months, allowing continuous exposure.

Root cause analysis

The incident is a classic security‑design flaw caused by missing default protection. The missing element is the X‑Robots‑Tag: noindex HTTP header or equivalent meta tag, which would tell crawlers not to index the page.

Impact: every user who employed the share function potentially exposed their data. The attack surface is the publicly reachable web page, and the vulnerability existed for several months.

Recommended defense layers

Technical layer: Add a default noindex tag or require explicit user confirmation before generating an indexable link.

Perception layer: Regularly monitor search‑engine indexing to detect accidental exposure.

Response layer: Establish an incident‑response workflow to quickly remediate reported leaks.

Design principle

Security researchers note that “one line of code can prevent this.” Applying a Privacy‑by‑Design approach—making “disallow search‑engine indexing” the default—would have avoided the breach.

Take‑away for blue‑team operators

Assume any data shared via Claude’s feature may already be compromised.

Immediately delete suspicious shared records.

Improve internal AI usage policies and prioritize data‑security controls.

AI tools are powerful, but they must be used with robust security practices to protect sensitive information.