Workflow Identity Hijacking: New AI Attack Vector Discovered

A new attack vector called 'workflow identity hijacking' has been identified, allowing attackers to bypass standard security controls in enterprise AI workflows. By sending benign requests through unauthenticated entry points, such as support emails or web forms, attackers can access sensitive data without needing to manipulate the AI model itself. This vulnerability arises from a design flaw in AI pipelines where the identity of the requester is decoupled from the permissions used to execute actions. Noma Labs reported that this flaw enables unauthorized access to privileged information, as the AI workflows execute actions using high-privilege service accounts. The attack was demonstrated by an example where an attacker received confidential sales data after sending a simple request. This issue is distinct from prompt injection attacks, which manipulate model instructions. The discovery emphasizes the need for improved authorization controls in AI systems.

Key Points: • Workflow identity hijacking allows unauthorized access to sensitive data via benign requests. • The attack exploits a design flaw in AI pipelines that decouples user identity from execution permissions. • Noma Labs emphasizes the need for enhanced security measures to mitigate this vulnerability.