Open models promise cheaper AI. But can they keep it secure?

Experts say the future of AI lies in combining open and closed models, both of which play important roles in cybersecurity.​

Mazin Gilbert, an executive director of the Agentic AI Foundation, which is part of the Linux Foundation, spoke during the Web Summit in Lisbon, one of the largest tech conferences globally.

Gilbert addressed concerns that open models come with cybersecurity risks. Such concerns were recently raised by Guillaume Princen, Anthropic’s head of international, who said that open-weight models carry “more risk by definition.”

“There are vulnerabilities, whether the model is open or closed, because we are going into the new phase of very advanced, sophisticated, intelligent models that we didn’t have in the past,” Gilbert said.

He added that while vulnerabilities in open models are visible to everyone, being able to see these problems also allows them to be fixed.

Regardless of the type of model used, Gilbert emphasizes that it's the business's responsibility to ensure they're secure.

“When you’re deploying your system, which the model is part of, you have to ensure you've gone through the security, you've invested in the maintainers, and ensure that your vulnerabilities are addressed,” Gilbert says.

With open-source models, their weights, the source code, and full training data are released, providing full transparency.

Meanwhile, more commonly used open-weight models keep their training data and training code secret. However, as they run on users’ hardware, they are touted for cutting costs, as they don’t come with a surprise token bill.

Philip Rathle, CTO at Neo4j, a global graph intelligence company, says open models can play an important part in cyber defenses.

He pointed out that Hugging Face was defending itself against the rogue OpenAI agents: the company used open-source models because they couldn’t access the frontier of sufficiently capable models.

OpenAI agents escaped the testing environment and hacked into the real-world company, Hugging Face, in the summer of 2026, sending shockwaves globally, as both businesses and governments are increasingly concerned with autonomous AI-related cybersecurity risks.

Open-weight models to cut AI costs

Interest in open models is growing, with nearly 6 in 10 senior executives at large companies planning to increase their use of cheaper open-weight ​AI models over the next year, according to a recent survey by Capgemini.

The transition is partly due to soaring costs of AI. European neobank Revolut, for example, saw its coding costs going down by 90% after the company moved 2,000 of its engineers from Anthropic’s Claude and OpenAI’s Codex to Chinese open-weight models.

German software company SAP SE recently removed some of the most expensive AI models from products ​, including its internal consulting assistant, and replaced them with lower-cost ​open-weight alternatives.

Gilbert says having more choice and not being locked into one model or provider will help drive down AI costs, including token costs and task completion costs.

However, he warns that having the best open-weight model doesn’t necessarily ensure having the best AI agents or the best system, because they are built on open data.

In July 2026, the open letter signed by major tech companies, including Microsoft, Mistral AI, Palantir, and Replit, called for the US to build “a strong, open ecosystem that diffuses into every sector.”

The companies argue that open-weight models are an important part of the system because they make “advanced AI more accessible, adaptable, and widely available.”

The letter reads, “Open weights also strengthen competition, and competition is what keeps the gains of AI broadly shared rather than concentrated in a few hands.”