Follow us on Google News
As businesses increasingly turn to cheaper open-weight artificial intelligence models, cybersecurity experts have warned that neither open nor proprietary AI systems are inherently secure. They say organisations must take responsibility for identifying vulnerabilities, maintaining safeguards and protecting the systems in which these models operate.
Speaking at the Web Summit in Lisbon, Mazin Gilbert, executive director of the Agentic AI Foundation under the Linux Foundation, challenged the assumption that open-weight models necessarily present greater security risks than closed alternatives.
Gilbert said vulnerabilities exist in both categories as AI systems become more sophisticated. He argued that the visibility of weaknesses in open models could also help security researchers identify and address them.
FCRF’s Chief AI Officer Certification Returns on November 14 to Prepare Future AI Leaders
Open Models Expose Weaknesses but Also Allow Faster Scrutiny
The debate has intensified as businesses increasingly consider open-weight models that can be deployed on their own infrastructure.
Unlike proprietary systems, open-weight models generally make their trained parameters available for use and modification, although their underlying training data and development processes may remain undisclosed.
This flexibility offers organisations greater control but also places more responsibility on them to maintain security.
Gilbert said the visibility of vulnerabilities in open models should not automatically be considered a disadvantage. Greater transparency can allow developers and security teams to examine weaknesses and develop fixes.
However, he stressed that choosing an open or closed model does not eliminate the need for proper security assessments.
Businesses deploying AI systems must ensure vulnerabilities are addressed, security investments are maintained and the wider applications built around these models are adequately protected.
AI Security Depends on the Entire System, Experts Say
Philip Rathle, chief technology officer at graph intelligence company Neo4j, also highlighted the potential role of open models in defending against cyber threats.
Speaking at the same event, Rathle referred to the use of open-source models by Hugging Face during its response to an incident involving autonomous OpenAI agents.
His remarks underlined the potential value of accessible AI models for organisations that may not have access to the most advanced proprietary systems.
Gilbert also cautioned that selecting a powerful open-weight model does not necessarily produce a secure or effective AI agent.
The performance and security of an AI application depend on how the model is integrated with other software, data sources and operational systems.
His comments place responsibility on organisations to examine the complete AI deployment rather than relying solely on the security characteristics of an individual model.
Cost Savings Drive Adoption as Security Responsibilities Grow
The security debate comes as companies seek alternatives to expensive proprietary AI services.
European digital bank Revolut reportedly reduced coding-related AI costs by 90% after moving approximately 2,000 engineers from proprietary coding assistants to Chinese open-weight models.
The shift illustrates the financial incentives encouraging businesses to deploy AI models with greater control over infrastructure and operating costs.
However, the savings also bring additional responsibilities for companies managing their own systems.
Gilbert argued that the availability of different models could reduce dependence on individual providers and encourage competition, but model choice alone would not determine the security or reliability of an AI application.
The discussion reflects a growing distinction between the affordability of AI technology and the resources needed to operate it securely.
For businesses adopting open-weight systems, the central challenge is ensuring that lower costs do not come at the expense of vulnerability management, maintenance and effective security controls.
About the author — Suvedita Nath is a science student with a growing interest in cybercrime and digital safety. She writes on online activity, cyber threats, and technology-driven risks. Her work focuses on clarity, accuracy, and public awareness.
