CyberSecurity SEE

Shadow AI: Agents Circumventing Data Governance Programs

Shadow AI: Agents Circumventing Data Governance Programs

The Rising Threat of AI in Data Security: An Unmanaged Challenge

In the evolving landscape of technology, the advent of artificial intelligence (AI) has sparked a multitude of developments within organizations, particularly in how data is managed and accessed. Among these innovations, one striking scenario involves a sales team coding an AI agent designed to automatically enhance and update client records across systems. In another instance, a developer generates an API key to gain access to valuable customer data. However, there’s also the common occurrence of a developer forgetting to deactivate a testing sandbox after experimenting with an AI agent. These instances are not mere technical hiccups; they represent severe implications for data security within organizations.

Conversations with numerous security professionals reveal a consensus that while the challenge of managing data security is not novel, the stakes have escalated. The surge in cloud adoption and Software as a Service (SaaS) has ushered in similar data governance and sprawl issues that were prevalent two decades ago. Nonetheless, experts assert that the current cloud-driven environment, accentuated by AI technologies, creates a more complex and perilous terrain at an unprecedented speed.

Mandar Chaudhari, a full-stack geospatial engineer at Land IQ and a researcher at George Mason University, voices concerns regarding the breadth of these challenges. “This problem existed long before AI,” he states. Developers have routinely created temporary scripts and test applications, yet the difficulty compounds when AI tools are tethered to a code repository or database. A significant concern arises when a developer inadvertently connects a coding agent to a testing environment. If that environment possesses broader credentials than necessary, the AI agent could exploit these permissions to traverse and manipulate data across the organization. Left unmonitored or abandoned, such an agent presents a sanctioned entry point for malicious actors, thereby circumventing conventional hacking methods.

Traditional access management solutions like privilege access management (PAM), data loss prevention (DLP), or next-generation firewalls (NGFW) face challenges in curbing this unauthorized access. Chaudhari elaborates that while API keys do not automatically bypass these security measures, their effectiveness largely depends on system configurations. He notes, “An API request using valid credentials can masquerade as normal application activity,” complicating detection efforts.

Adding to the complexity of data security is the issue of unmanaged third-party integrations. Ryan Reck, founder of Reck Advisory, highlights that SaaS-to-SaaS agents often operate seamlessly in the background, complicating the discovery and control of unsanctioned integrations. AI tools can demand extensive access to mailboxes, shared drives, and customer relationship management (CRM) systems. With self-service consent mechanisms, employees can readily approve these connections without the necessary review from procurement or IT departments.

Furthermore, AI initiatives frequently leave residual data such as test datasets and logs that do not fall under established data governance frameworks. Reck explains that this scenario mirrors issues of shadow IT and cloud sprawl, while also presenting new avenues for connecting systems that risk proliferating enterprise data copies.

The implications of these risks are not merely theoretical. A notable incident occurred in April when a data breach at Vercel initiated from a compromised third-party AI tool, Context.ai. This breach enabled the unauthorized access of an employee’s Google Workspace account, eventually leading to lateral moves within the Vercel network, which resulted in critical data being encrypted.

In contrast to traditional cloud developer integrations that typically establish one-on-one connections, AI integrations often link multiple systems or even connect to model context protocol (MCP) servers. This multifaceted connectivity presents a risk whereby forgetting a single integration can expose sensitive data across numerous repositories.

Forgotten developer sandboxes also pose a persistent security challenge. As emphasized by Mudita Khurana, security engineer at Airbnb, cloud-based development platforms allow for the unchecked duplication of data into environments unknown to IT and security teams. AI exacerbates this dilemma, as developers can link third-party large language models (LLMs) directly to data repositories using API keys. Khurana underscores that while an intentional data breach can occur via an unapproved AI tool, the unintentional leakage often poses an even greater risk. Instances may arise where an initial integration was sanctioned for an experimental purpose, yet its authorization remains in effect long after the project has ceased.

The issue of API keys further complicates matters; many do not have expiration dates, allowing for continued access to sensitive information long after the original project has been abandoned. If exposed, these keys could grant unauthorized users entry into critical systems without necessitating a traditional hacking approach.

Bapuji Koraganti, Principal Engineer at Comcast and an independent identity researcher, points to patterns of AI-related risks centering on credentials and delegated grants. He stresses the importance of recognizing that while the trends of unmanaged assets are not new, the dynamic of AI introduces an alarming level of authorized but unaccountable access. A solitary click on ‘Allow’ can initiate an autonomous process that exploits broad permissions indefinitely.

Koraganti identifies two crucial risks that require greater scrutiny: the revocation gap, where a compromised integration may continue to have access due to cached tokens, and permission flattening, which allows data access through a broadly privileged service account. Such vulnerabilities become significantly concerning as AI constructs can bypass conventional safeguards, hence leading to an overwhelming data exposure threat.

While these challenges echo the historical vulnerabilities seen during the cloud era, the rapid pace of AI development complicates the ability of cybersecurity and data governance teams to respond effectively. Reck aptly captures the essence of this predicament: “A back door into company data can start at the front door, with an employee clicking ‘Allow.’ Someone needs to own that connection long after the trial ends.” As organizations continue to navigate this intricate landscape, fostering stringent controls and robust oversight is essential to mitigate the widening chasm of data security risks exacerbated by AI.

Source link

Exit mobile version