Research Highlights Identity Governance Risks in AI Agents

New research describes a technique dubbed “GhostJacking,” in which attackers may exploit security alerts and blocked events to influence the behavior of AI agents. The findings point to weaknesses in how agent identities, permissions and security signals are governed.

AI agents increasingly operate with access to applications, data and automated workflows. As a result, the systems that monitor their activity can become an important part of the security boundary. According to the research, an attacker may attempt to use alerts or denied actions as inputs that alter an agent’s decisions, potentially creating an opportunity to manipulate or hijack its operation.

Why the issue matters

Traditional identity controls are generally designed around human users and long-running service accounts. AI agents, by contrast, can make decisions dynamically, invoke multiple tools and react to events in real time. If their identities and permissions are not tightly managed, an agent could receive more access than necessary or respond unpredictably to hostile signals.

The research underscores that blocked activity should not be treated as harmless background noise. Security events may also be processed by the agent itself, by connected orchestration systems or by other automated components. Without clear separation between trusted instructions and untrusted event data, attackers may be able to influence an agent’s context or actions.

Security considerations

• Assign each agent a distinct, tightly scoped identity and limit access to only the tools and data it needs.
• Separate security telemetry from instructions that agents are authorized to follow.
• Review how alerts, denied requests and other external inputs are validated before reaching an agent.
• Monitor agent activity for unexpected permission use, repeated blocked actions or abrupt changes in behavior.
• Maintain human approval for sensitive operations, particularly those involving identity, access or irreversible changes.

The findings add to growing concerns that securing AI agents requires more than model safeguards. Organizations must also apply identity governance, least-privilege controls and continuous monitoring to the agents and the systems that direct them.