Enterprises Face Governance Risks as Open-Weight AI Adoption Surges
Share
Enterprise adoption of open-weight artificial intelligence (AI) models has seen a sharp increase in late 2026, with approximately 42% of surveyed organisations now running these models in production environments.
While open-weight models offer lower costs and improved data privacy, security researchers are warning of a significant “governance gap” as companies fail to implement necessary controls for these systems.
Research from Cato Networks indicates that the shift toward open-weight models—where the learned parameters or “weights” are available to download—requires organisations to take on security responsibilities typically managed by AI service providers.
The Security Trade-off of Model Ownership
When an organisation uses a hosted AI service, the provider manages the infrastructure, safety controls, and updates. In contrast, running an open-weight model on private infrastructure gives the enterprise greater control but introduces new attack surfaces.
Security analysts have identified several critical risks associated with the DIY approach to AI deployment, including the potential for Trojan horses. Some model files can execute malicious code when loaded into a system, while others may contain hidden backdoors that can be triggered by specific inputs.
The provenance of models is also a growing concern. Models hosted on public repositories may be modified by third parties without clear documentation, potentially introducing weakened safety rules or malicious modifications that are difficult to detect without rigorous testing.
Operational Challenges and Vulnerabilities
Unlike software-as-a-service (SaaS) platforms, open-weight models do not update automatically. Security teams must actively track developer releases, test new versions, and manually apply updates to ensure they are protected against newly discovered vulnerabilities.
Furthermore, running these models increases the total IT attack surface. Organisations must secure the physical servers, graphics processing units (GPUs), and the specific training data sets used to fine-tune the models for internal use.
Establishing AI Governance Frameworks
To mitigate these risks, security experts recommend that governance should be integrated into the deployment phase. Organisations are encouraged to assign clear ownership for every model in use, ensuring a specific team is responsible for approving its deployment and responding to security incidents.
Key recommendations for securing open-weight AI include:
- Centralised Repositories: Create approved internal repositories to fetch and store models, rather than allowing teams to download directly from unvetted public sources.
- Integrity Verification: Record file hashes for every model and its subsequent fine-tunes to detect unauthorised alterations.
- Adversarial Testing: Test models as an attacker might, attempting to bypass safety rules or elicit unexpected behaviour before moving the system to production.
- Isolation: Run higher-risk models, such as those used for reading source code, in isolated environments with limited access to sensitive credentials or keys.
As AI agents and autonomous systems become more prevalent, the need for robust oversight becomes more urgent. Security leaders suggest that open-weight models should be scrutinised with the same rigour as any other piece of critical enterprise software brought in-house.



