Open-source AI and its implications for international security

17 November 2026
Open-source AI and its implications for international security

Open-source artificial intelligence (AI) remains an under-addressed aspect of the developments in the field. The risks associated with its proliferation among malicious non-State actors are growing rapidly, while discussions and measures to address them have yet to keep pace with innovation and technological capability diffusion.

Ongoing research by UNIDIR’s Centre of Excellence on AI, Peace and Security aims to unpack the security risks of open-source AI, as well as propose key policy recommendations for tackling this growing challenge. The resulting study is part of the Institute’s work on countering the proliferation of AI.

What is open-source AI?

While the concept of open-source has been established in software for decades, its application to AI is relatively novel, raising new concerns in the context of international security. To describe a software license as open-source means that its source code – the human-readable instructions from which a programme is built – is openly available for anyone to use, study, modify and redistribute.

Although the term is not directly applicable to AI models, whose components are more complex than simple code, the degree to which they can be accessed varies. Defining AI models in terms of an open-closed binary would be misleading, and whether a given model can be categorized as open-source requires a nuanced discussion. Still, AI models can be broadly categorized as:

  • fully closed, contained strictly within creator companies,
  • cloud-based, run by private companies and often accessed through a subscription (currently how most people use AI),
  • open-weight, which anyone can download, run and modify, but which retain specific elements (e.g. training data, code) private, and
  • fully open-source, whose components are available to everyone.

The most prevalent type of open AI models are open-weights models. While not fully open, these models make their weights – the key parameters that define the responses produced by a model – available for access and modification.

Despite being technically inaccurate, the terms open-source and open-weight are often used interchangeably. Still, the two can essentially be viewed as equivalent in terms of the risks associated with their proliferation among malicious actors.

Once released, an open AI model can be neither recalled nor regulated, as they work locally, without depending on an external entity (like a private company) for computing, running or updating the model, and gathering user data.

Among the benefits, hardware requirements are usually much lower for open models. For example, a high-end laptop is sufficient to run a local AI model for automated social media content generation, basic intelligence analysis and logistical support for a small-scale organization. In addition, open models allow researchers, governments and private entities with sensitive data like banks, to maintain sovereignty over AI models and adjust them to users’ own needs and purposes. Open-source AI models also grant equal opportunities and reduce dependencies for developing countries.

However, open-source AI is a perfect example of dual-use technology. While providing ample benefits, it simultaneously gives malicious actors new avenues for AI misuse and weaponization.

What are the security risks associated with open-source AI?

Open models are open to everyone – researchers and malicious actors alike – and offer both the same privacy and degree of sovereignty. While companies that control access to closed models can monitor users’ requests and deny access to suspected malicious actors, open models have barely any measures beyond built-in guardrails to ensure their safe use. These guardrails, however, are easily removed or bypassed within days of a model’s release. With guardrails removed, open models grant full access to all the knowledge they contain, from gardening tips to explosives manufacturing.

Models with removed guardrails, also known as ‘uncensored’ or ‘abliterated’, can significantly increase the capabilities of malicious non-State actors. For instance, they can help malicious actors plan attacks, manufacture or acquire weapons, manage logistics for large-scale operations, recruit members and gather intelligence, among other activities that can support violent extremist or terrorist groups.

Cases of terrorist groups’ use of closed models for logistics and arms and explosives manufacturing have also been documented, as well as similar cases involving extremists and terrorists acting alone. While AI companies can detect and flag these actors’ use of closed AI models and suspend or ban accounts on suspicion of violent intent, open models are not subject to the same oversight.

Once downloaded and uncensored, an open model cannot refuse a request or alert anyone of an imminent threat.

How can the risks be addressed?

To inform policy responses, forthcoming work by UNIDIR’s Centre of Excellence on AI, Peace and Security will contribute to filling an important gap in current discussions about AI risks. The research will examine key aspects of open-source AI, its use cases, and how regulatory frameworks may respond to the challenges it presents.

A key challenge is the dual-use nature of this technology, which complicates the regulation of training data.

The research will explore possible responses, ranging from technical measures, such as safeguards against the removal or guardrails and identity verification, to policy and multilateral cooperation to counter proliferation and misuse.

The security implications of open models are dire, with risks evolving and multiplying rapidly. Since the release of open models is irreversible, each new, more capable, unmoderated model presents new challenges. However, to address their potential harm, measured and coordinated responses are needed that can help prevent malicious use while maintaining fair and equal access to the benefits granted by open models.