Research highlights how over-permissioned AI agents can become insider threats
Unit 42, Palo Alto Networks’ threat intelligence team, has uncovered a set of security risks in Google Cloud’s Vertex AI platform that could allow malicious or compromised AI agents to access sensitive data and cloud resources beyond their intended scope.
The research focuses on Vertex AI Agent Engine, a platform used to build and deploy autonomous AI agents capable of interacting with enterprise systems, data and services.
At a high level, Unit 42 demonstrated how an attacker could create a seemingly legitimate AI agent that secretly extracts its own credentials and uses them to gain broader access within a cloud environment. This behavior effectively turns the agent into a “double agent,”operating as both a trusted tool and a potential insider threat.
Overview of the Attack Mechanism
The issue stems from how permissions are assigned to AI agents by default. Unit 42 found that service accounts linked to deployed agents were granted overly broad permissions, enabling access to resources beyond what was strictly required. By exploiting this, researchers were able to extract credentials and use them to:
- Access data stored in cloud storage within the customer environment
- Retrieve sensitive deployment information and configurations
- Gain visibility into restricted internal components supporting the AI platform
Importantly, this was not a single vulnerability, but rather a chain of misconfigurations and design gaps that, when combined, expanded the agent’s effective access.
Broader Security Implications
As organizations increasingly adopt AI agents to automate workflows and decision-making, these systems are being granted high levels of trust and access.
This research highlights a critical shift in the threat landscape:
- AI agents can act autonomously, often without continuous human oversight
- If compromised, they behave like trusted insiders, not external attackers
- Over-permissioned agents can significantly expand the attack surface
The findings underscore the risks of deploying AI systems without strict adherence to the principle of least privilege.
Mitigation and Industry Response
Palo Alto Networks responsibly disclosed the findings to Google. In response, Google updated its documentation to provide greater clarity on how Vertex AI uses service accounts and permissions.
The research highlights the need for organizations to institutionalize rigorous AI security reviews as part of their deployment lifecycle. This includes enforcing least-privilege access through the use of dedicated, custom service accounts such as Bring Your Own Service Account (BYOSA), carefully validating permission boundaries, and restricting OAuth scopes to prevent unnecessary access. It also underscores the importance of treating AI agent deployment with the same level of scrutiny as production code, including conducting thorough security reviews prior to deployment.
As AI agents become more autonomous, ensuring tighter control over their permissions and behavior will be critical to minimizing risk. Solutions such as Prisma AIRS, Cortex AI-SPM, and Cortex Cloud Identity Security can support organizations in addressing this emerging AI security gap.
The findings point to a broader architectural challenge: as AI systems become more deeply integrated into enterprise infrastructure, security risks increasingly emerge from how components interact, rather than from isolated software flaws.
Even when individual systems function as intended, their combined behavior can introduce unintended exposure. As AI adoption accelerates, organizations will need to rethink how they manage trust, permissions and isolation; particularly for autonomous systems that can act on their behalf.
