Varonis Threat Labs built an AI agent called Pinchy on OpenClaw and tested it against phishing attacks in a controlled Google Workspace environment. Despite safety prompts, the agent forwarded AWS IAM keys, database passwords, SSH credentials, and a CRM export containing 247 enterprise customers and $1.28M in MRR to an attacker. The agent performed better against technical phishing (e.g., a malicious OAuth flow) but failed on socially-engineered requests framed as routine colleague tasks. Security experts attribute the failures to architectural flaws: agents mixing data and control channels, over-privileged access, missing human oversight, and absent runtime guardrails. Recommendations include treating AI agents as high-privilege identities, enforcing tool-level permissions, and requiring human review for sensitive data-sharing actions.
