AI Agents Pose New 'Synthetic Insider Threat' to Organizations

A new report by Israeli cybersecurity firm Above Security identifies 166 ways artificial intelligence agents can become an internal threat to organizations, a phenomenon they term 'synthetic insider threat.' Unlike traditional insider threats, which involved human employees with malicious intent, these AI agents operate with employee credentials and at machine speed, posing a significant and often undetected risk.

The report, developed by Nimr Qis, Head of AI Research, and Yonatan Malka, Head of Solutions at Above Security, highlights that AI agents can be manipulated or malfunction without malicious intent, leading to data breaches or unauthorized actions. Data from Verizon and Palo Alto Networks indicates a rapid adoption of AI by employees, with 45% regularly using AI tools on work devices, often through personal accounts, and a lack of robust logging by most organizations, making it difficult to track AI activity.

Above Security outlines four scenarios where AI agents can go rogue: a compromised internal AI assistant, an agent tricked by hidden instructions in documents, an AI tool with excessive permissions, a 'poisoned memory' where malicious instructions persist, and an agent misinterpreting commands literally. Warning signs include non-human activity rates, data exfiltration to unknown destinations, access beyond granted permissions, discrepancies between reported and actual outcomes, and a single identity operating from multiple locations simultaneously.

To mitigate these risks, Above Security recommends assigning unique, time-limited identities to AI agents, granting permissions strictly based on the task at hand, treating all content read by AI as data rather than commands, requiring human approval for irreversible actions, and maintaining immutable logs of all AI activities. They also suggest revoking an agent's permissions rather than immediately disabling the employee's account during an incident.

Meanwhile, Guy Kozliner, CEO of Rig Security, another Israeli cybersecurity firm that recently raised $12 million, emphasizes that traditional authentication methods are insufficient. He warns that AI agents, even when legitimately accessed by an employee, can exploit broad permissions to cause significant damage, potentially creating backdoors for attackers. Kozliner projects a dramatic increase in AI agents within organizations, with Fortune 500 companies expected to run over 150,000 AI agents by 2028, underscoring the urgent need for organizations to develop robust control mechanisms.