Google Vertex AI flaw exposes security risks with elevated privileges in service agents

Google Vertex AI flaw exposes security risks with elevated privileges in service agents
Google Vertex AI, cybersecurity, vulnerability, XM Cyber, artificial intelligence
A detailed security analysis identified two significant vulnerabilities in the Google Vertex AI artificial intelligence platform. The discovery, released by cybersecurity company XM Cyber, reveals that users with minimal permissions can escalate their privileges to access more powerful functions within the system. The flaws exploit default settings of so-called “service agents”, automated accounts that operate behind the scenes on the platform.
The central problem lies in the way the platform manages the permissions of these service accounts, which are created to facilitate the integrated operation of different AI tools. Segundo researchers, this architecture allows a malicious actor with basic view access to manipulate the system to obtain elevated privilege access tokens, opening doors for unauthorized activity. Essa breach represents a considerable risk for organizations that rely on the platform to develop and operate their machine learning models.
In an official response, Google stated that the observed behavior is in line with the platform’s original design, classifying it as intentional. Essa position highlights a growing debate in the technology industry about balancing operational convenience and implementing rigorous security controls in increasingly complex cloud environments.
Technical details of the vulnerabilities
The vulnerabilities exploit the way privileges are allocated to “service agents” on Google Cloud. Essas special accounts, managed directly by Google, are designed to interact with user resources and execute internal processes autonomously. Para ensure continuous operation and integration between services, they receive, by default, a broad set of permissions at the project level, often exceeding what is necessary for specific tasks.
Research by XM With this token in hand, the attacker assumes the identity of the service account and can perform actions that would be far beyond its original scope, transforming an internal management account into a powerful vector for privilege escalation.
The official position of Google Cloud
Following XM Cyber’s responsible disclosure of the failures, Google Cloud maintained its position that the services operate as planned. The company did not provide additional comment or announce plans to change the default setting. The current architecture prioritizes ease of use and rapid activation of AI functionalities, delegating to service agents the broad access necessary to orchestrate complex operations without constant manual intervention.
Security analysts argue that this approach, while convenient for developers, subordinates corporate governance models to the cloud provider’s architecture. Companies using the platform lose visibility and granular control over these “invisible” identities operating in their environment. Essa lack of transparency makes it difficult to apply the principle of least privilege, a fundamental pillar of cybersecurity.
This design philosophy places the responsibility for implementing additional controls squarely on customers. Sem With proactive configuration and monitoring, companies are exposed to risks that are not immediately apparent, believing that the provider’s default settings are inherently secure. The situation serves as a wake-up call to the need for rigorous due diligence when adopting managed AI services.
Magnified risks for insider threats
The weaknesses identified in the Vertex AI significantly amplify the danger posed by insider threats. A malicious employee or attacker who has compromised a low-privilege account can exploit this mechanic to move laterally across the network and access sensitive data or critical systems without being detected. Service agent manipulation is particularly insidious because it does not generate the typical security alerts associated with suspicious user account activity.
Traditional enterprise security tools, such as intrusion detection systems or identity management platforms, are rarely configured to monitor the behavior of provider-managed identities. From the logs’ point of view, the abuse of these accounts appears as legitimate, authorized operations of the platform itself, masking the malicious activity. Isso creates a dangerous blind spot for security teams.
AI environments often process and store large volumes of proprietary and sensitive data. The lack of adequate isolation between different projects and workloads means that the compromise of a single service agent can have a devastating impact. An attacker could, for example, access and exfiltrate training datasets, manipulate machine learning models to produce incorrect results, or use computing resources for illicit purposes, such as cryptocurrency mining.
Malicious actions may include monitoring unexpected queries against databases such as BigQuery, unauthorized access to data storage buckets, performing abnormal behavior in API sessions, or changing data pipeline orchestration settings. Cada one of these actions can cause significant financial and reputational damage to the affected organization.
Context of previous issues in Vertex AI
This is not the first time that the security of the Vertex AI platform has been questioned. In November 2024, researchers at Palo Alto Networks disclosed a different set of vulnerabilities that also allowed privilege escalation and exfiltration of AI models. On Naquela occasion, the failures were related to the execution of “custom jobs” and improper access to machine learning models that had been adjusted with proprietary customer data.
Google promptly responded to Palo Alto Networks’s disclosure and applied fixes to the specific issues identified. However, the recurrence of vulnerabilities linked to privilege management suggests the existence of structural design patterns on the platform. Granting excessive permissions by default to internal actors appears to be a fundamental engineering philosophy, which, while it speeds up development, creates fertile ground for exploitation by adversaries.
Implications according to OWASP Agentic Top 10
The vulnerabilities in Vertex AI directly align with emerging risks cataloged by the security community. In December 2025, the OWASP foundation, a global reference in application security, released the “Top 10 for Aplicações Agentic”, a guide dedicated to the specific risks of autonomous AI systems. Category ASI03, “Abuse of Identidade and The OWASP framework serves as a crucial tool for organizations to assess and mitigate the risks associated with the adoption of AI platforms, highlighting that traditional security threats take on new dimensions in autonomous ecosystems.
Recommendations from security experts
Given the risks, security professionals recommend that companies using Vertex AI immediately implement robust compensatory controls. It is crucial that Chief Information Security Officers (CISOs) and their teams conduct a thorough audit of all service identities associated with their AI workloads, treating these accounts not as passive components of the platform, but as privileged users that require constant monitoring. Implementing alerts focused on usage patterns that indicate improper manipulation is an essential measure to detect suspicious activities that would go unnoticed by conventional systems.







