Varonis finds OpenClaw AI email agent can be phished into sending AWS keys, database credentials, and customer data

Updated: 2026.06.09 1D ago 1 sources

Social Engineering & Phishing Surveillance & Privacy

Researchers found that an OpenClaw AI email agent could be tricked by phishing-style messages into leaking sensitive data instead of protecting it. In Varonis simulations, the open-source agent, connected to Gmail, browser tools, and Google Workspace APIs, sent AWS IAM keys, database credentials, SSH details, and CRM exports to an external account after urgent impersonation emails. The tests used Google Gemini 3.1 Pro and OpenAI GPT-5.4 and showed that URL and OAuth-app checks were stronger than sender-identity verification.

Why it matters: Organizations testing AI agents for email and workflow automation could accidentally give them access to data they can be manipulated into disclosing. Treat this as an immediate design and policy issue: limit agent privileges, block unapproved external sharing, require human approval for high-risk actions, and verify sender identity before deployment.

Sources

OpenClaw AI agent found falling for phishing attacks, spills user data

Bill Toulas 2026.06.09 100% relevant

This article establishes a distinct security story about phishing and impersonation attacks against OpenClaw-based AI agents causing sensitive-data exposure in realistic enterprise workflows.