Identity-Based AI Attack Threatens Security of Enterprise Data
Dark Reading, Wednesday, September 9th, 2026
Noma Labs' 'workflow identity hijacking' lets attackers abuse privileged AI workflows through unauthenticated inputs to steal data.
Researchers at Noma Labs identified 'workflow identity hijacking,' an authorization design flaw in enterprise AI pipelines that decouple the identity of whoever triggers a workflow from the privileges used to execute it.
Attackers send benign-looking requests through unauthenticated entry points such as support inboxes, GitHub issues or web forms, and the workflow executes them with high-privilege service accounts, for example emailing back a finance director's messages.
Experts stress this is an identity issue rather than prompt injection. Recommended defenses include scoped, short-lived delegation tokens, authorization checkpoints treating LLM output as untrusted, separating data retrieval from external responses, and deception assets.