Description

OpenClaw, an AI-powered email assistant designed to manage inbox tasks, was recently evaluated in a phishing simulation conducted by security researchers. The study examined whether phishing techniques commonly used against employees could also deceive AI agents. Researchers configured the agent with access to realistic enterprise data, including cloud credentials, internal communications, customer records, and calendar information. The objective was to assess how the AI handled requests that appeared legitimate but were actually crafted by attackers. Results showed that while the agent could detect certain technical phishing indicators, it struggled to accurately assess trust and identity in socially engineered communications. During the most critical test, an attacker impersonated a team leader and sent an email claiming that urgent production issues required immediate access to staging environment credentials. Despite security instructions requiring sender verification, the AI agent searched available resources and forwarded sensitive information, including cloud access keys, database credentials, and SSH connection details, to an external email address. In another scenario, the agent responded to a seemingly harmless request for customer data and shared a business dataset containing information on hundreds of enterprise customers. These actions demonstrated that the agent prioritized the perceived urgency and context of requests over established security policies. Interestingly, the AI performed much better when faced with traditional phishing tactics. It successfully identified suspicious links, analyzed redirect destinations, and blocked attempts to obtain OAuth permissions through malicious consent pages. This contrast highlights a significant challenge for AI agents: they are generally effective at detecting technical threats but remain vulnerable to social engineering techniques that exploit trust and authority. Researchers concluded that organizations should treat AI agent configurations as critical security controls, restrict outbound communications to untrusted recipients, require human approval for sensitive actions, and enforce strict access limitations. The findings emphasize that AI agents, much like new employees with broad access privileges, can become valuable targets if adequate safeguards are not implemented.