CISO Concerns Rise As [Company Name] Adopts On-Device AI

For the last 18 months, Chief Information Security Officers (CISOs) have focused on controlling browser access to manage generative AI risks. However, a new challenge has emerged with employees running AI models locally on their devices, bypassing traditional network controls. This shift, often referred to as “Shadow AI 2.0” or “bring your own model” (BYOM), presents significant security blind spots for enterprises.

## The Rise of Local Inference

The ability to run large language models (LLMs) locally on laptops has become increasingly practical. This is due to advancements in consumer-grade hardware, such as MacBook Pros with 64GB memory, and the mainstream adoption of model quantization techniques. These developments allow high-performance models to run efficiently on personal devices without requiring cloud-based infrastructure.

The distribution of open-weight models has also become seamless, enabling engineers to download and operate these models with minimal effort. As a result, sensitive workflows can now be executed offline, leaving no network signature or audit trail. This creates a scenario where traditional data loss prevention measures are ineffective, as they are designed to monitor data leaving the network, not local device activity.

## Security Risks and Implications

The shift to local inference introduces several new risks. Integrity risk arises when unvetted models are used to generate code or make decisions, potentially degrading security posture without detection. Compliance risk is another concern, as many models come with complex licensing terms that may be violated when used locally without oversight. Finally, provenance risk involves the accumulation of unverified model artifacts on endpoints, which could contain malicious payloads.

These risks highlight the need for enterprises to adapt their security strategies. Traditional network controls are insufficient for managing local model usage, necessitating a focus on endpoint governance. This includes inventorying model artifacts, monitoring device activity, and ensuring compliance with licensing terms.

## Adapting to the New Landscape

To mitigate the risks associated with BYOM, companies should implement endpoint-aware controls and provide a curated internal model hub. This hub would offer approved models, verified licenses, and guidance for safe usage, reducing the need for employees to seek external, potentially risky alternatives. Updating policy language to explicitly cover local model usage is also crucial, ensuring clear guidelines for employees.

As AI activity increasingly shifts to endpoints, organizations must focus on controlling artifacts and ensuring compliance at the device level. This approach will help maintain security without stifling productivity, as the perimeter of AI governance moves from the cloud back to individual devices.