In a recent cybersecurity test, Varonis researchers exposed vulnerabilities in the OpenClaw AI agent, known as Pinchy, which was tricked into phishing attacks, leading to potential user data compromise. The test highlighted the AI's inability to handle identity verification under urgent scenarios, despite its effectiveness in blocking malicious links and OAuth applications.
The Experiment Setup
Varonis connected the OpenClaw agent to a Gmail inbox, browser tools, and Google Workspace APIs, populating it with fake internal company data, including AWS and database credentials. Two configurations were tested: a generic mode with standard productivity instructions and a strict mode designed to be aware of phishing scams.
AI's Mixed Performance
The AI's performance was mixed. When attackers impersonated a team lead requesting access to the staging environment, Pinchy granted it. Similarly, it complied with a request for a customer export under the guise of remote work. However, the AI successfully blocked a phishing link in a fake gift card email and denied access to a malicious OAuth application disguised as a timesheet platform.
"Both Generic and Strict profiles failed because the verification step still collapsed when the request appeared operationally urgent," Varonis reported.
Model Comparisons
Varonis tested two models: Gemini 3.1 Pro and GPT-5.4. The researchers noted that Gemini showed a greater willingness to interact, while GPT was more cautious. This suggests a need for enforced identity verification before AI agents proceed with sensitive actions.
Implications for Cybersecurity
The findings underscore the importance of robust identity verification processes in AI systems, especially in scenarios where requests appear urgent. For CTOs and technology leaders, this highlights the need to evaluate AI tools not just for their ability to detect malicious content but also for their capacity to verify identities effectively.
| Model | Interaction Willingness | Caution Level |
|---|---|---|
| Gemini 3.1 Pro | High | Low |
| GPT-5.4 | Low | High |
The study by Varonis serves as a critical reminder of the evolving challenges in cybersecurity, particularly as AI becomes more integrated into business operations.
Sources:
