Security breach in Google Vertex AI allows escalation of privileges through service agents

An investigation conducted by cybersecurity firm XM Cyber revealed the existence of two significant vulnerabilities in the Google Vertex AI artificial intelligence platform. The flaws allow users with minimal permissions to escalate their privileges, accessing system functions with much higher authority than originally designated. The central problem lies in the tool’s default settings, which grant broad powers to so-called “service agents”, automated accounts that operate in the background.
According to the researchers, exploiting these loopholes turns low-risk access, such as that of a simple viewer, into a potential vector for complex insider attacks. In response to the responsible disclosure of the flaws, Google reported that the observed behavior is in line with the platform’s original design, a position that sparks debate about the balance between functionality and security in cloud-managed AI services.
The alert reinforces experts’ growing concern about the security of artificial intelligence environments, where the complexity of internal operations can mask critical risks. The discovery highlights the need for more rigorous security governance on the part of companies that adopt these technologies, in order to prevent default configurations from becoming gateways for internal or external threats.
The mechanism behind privilege exploitation
The vulnerabilities identified by XM Cyber are intrinsically linked to the way in which Google Cloud manages the identities of its “service agents”. Essas are special service accounts, controlled by the platform itself, that act on the user’s behalf to perform automated tasks and integrate different resources. Para ensures that AI services run seamlessly and at maximum efficiency, Vertex AI grants these agents very broad project-level permissions. The flaw lies in the ability of an attacker, even with a limited access profile such as “Viewer”, to manipulate the system to obtain authentication tokens from these privileged agents under certain conditions. Once in possession of these tokens, the attacker can effectively assume the identity of the service agent and perform actions that would be far beyond their original permissions, such as accessing sensitive data, modifying configurations or interacting with other Google Cloud services, all of this appearing to be a legitimate operation of the platform.
Google Cloud’s official response
After being notified by XM Cyber, Google Cloud reviewed the report and classified the behavior as intentional and in line with the Vertex AI design. The company has not issued additional public comments, maintaining its position that the services operate as expected.
This approach prioritizes convenience and agility, allowing developers to enable and use advanced AI capabilities immediately, without the need for complex permissions configurations. Service agents operate with broad access to ensure that integrations between different tools, such as BigQuery and Cloud Storage, work smoothly.
However, security analysts point out that this “security by default” philosophy can subordinate corporate governance models to the provider’s architecture. Client companies lose visibility and granular control over crucial internal components, blindly trusting vendor-implemented security, which can create significant protection gaps.
Implications for corporate security
The main consequence of these vulnerabilities is the increase in the risk associated with internal threats. A malicious employee or compromised account with least privileges could exploit this weakness to move laterally within the cloud environment, gaining access to sensitive information or sabotaging critical operations. Tampering with service agents is particularly dangerous because it generally does not trigger traditional security alerts, which are calibrated to detect suspicious activity from human user accounts.
Enterprise security monitoring tools are rarely configured to monitor the behavior of provider-managed identities, causing abuse of these accounts to be masked as a routine operation of the platform. In AI environments, which often process large volumes of sensitive data, the lack of adequate isolation and specific monitoring drastically increases the impact radius of an eventual compromise. Atividades how unexpected database queries, unauthorized access to storage repositories, and changes to orchestration configurations can go unnoticed.
History of vulnerabilities on the platform
This is not the first time that the security architecture of Vertex AI has been questioned. In November 2024, researchers from Palo Alto Networks disclosed other vulnerabilities on the same platform that also allowed privilege escalation, but through “custom jobs” and the exfiltration of machine learning models.
At that time, Google implemented fixes for the issues highlighted, which involved custom training pipelines and improper access to fine-tuned AI models. The recurrence of failures related to privilege management suggests a structural pattern in the design of managed AI platforms, which, by default, tend to grant elevated permissions to internal agents to simplify the user experience.
Alignment with OWASP guidelines
The risks highlighted in Vertex AI are in full compliance with the new security guidelines for AI systems. In December 2025, the OWASP foundation, a global reference in application security, released the “Top 10 for agentic applications”, a list of the main risks in autonomous systems.
The ASI03 category, “Abuse of Identidade and The framework serves as an essential guide for organizations to assess the risks inherent in platforms like Vertex AI and implement appropriate defenses.
Mitigation measures and security recommendations
Cloud security experts recommend that companies using Vertex AI not wait for platform changes and implement compensatory controls immediately. The first action should be a thorough audit of all service identities associated with AI workloads.
It is essential to treat service agents as highly privileged accounts, applying specific monitoring to detect usage patterns that may indicate manipulation or abuse. Isso includes the configuration of alerts for anomalous behaviors that deviate from the expected profile of an automated operation.
Other important measures involve reducing the scope of authentication between different cloud components where possible and introducing stronger security barriers to protect the most critical assets. Segmenting sensitive resources into separate projects or virtual networks can limit an attacker’s reach.
Periodically reviewing the default permissions granted by the platform is an essential security hygiene practice. While these actions may increase operational complexity and costs, they are crucial to limiting the blast radius of an incident and ensuring effective control over invisible identities operating in the environment.
The dilemma of shared responsibility
The accelerated evolution of AI tools requires constant adaptation in defense strategies. Provedores and customers share responsibility for the security of internal components, and blind trust in default configurations from managed providers exposes organizations to critical security gaps in modern cloud environments.







