Open Models Have Real Benefits. Do They Outweigh The Risks?

TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

An October 13, 2026, MIT Technology Review report examines the trade-offs in making powerful AI models downloadable. Open models can support research, local use and competition with major AI companies, but their availability can also make it harder to prevent misuse; the report describes the scale of that risk and the best policy response as contested.

A debate over whether powerful AI models should be downloadable or kept behind company-controlled services has sharpened after Anthropic warned of cyberattack risks from GLM-5.3, a model released in August by Chinese company Z.ai. In an October 13 analysis, MIT Technology Review examined the competing concerns: open models can widen access and support research, while their availability may make harmful uses harder to control.

The central distinction is between models people can access only through a service or application programming interface and models whose weights can be downloaded. Those weights encode a model’s learned parameters. The report notes that “open model” is often used broadly; “open-weight” is a more precise term when weights are public but other information, such as training details, is not.

Downloadable models can be run on a user’s own hardware and, with sufficient expertise and resources, altered through retraining or fine-tuning. Researchers can inspect more than a model’s prompt-and-response behavior, combine models, and experiment with efficiency. Businesses may use them locally to handle sensitive or proprietary information. A Mozilla report cited by MIT Technology Review found open models typically cost less than closed models, though the article does not provide a specific price comparison.

The report says closed models still led the Artificial Analysis LLM leaderboard at publication, with the top eight models all closed, while open models were described as only months behind. It presents the risk case through Anthropic’s assessment of GLM-5.3; that assessment is a company warning, not independent proof that the model has been used in a cyberattack.

Who Controls AI Infrastructure

The debate affects more than developers choosing a tool. If a small number of companies control access to the most capable systems, they can shape who uses them, on what terms, and for which purposes. Open models offer an alternative by allowing more people and organizations to run systems without depending entirely on a handful of providers.

That wider access also complicates safeguards. A company operating a closed service can set usage rules and restrict access, while a downloadable model can be copied or modified beyond its publisher’s direct oversight. The competing interests are substantial: the report notes that Nvidia CEO Jensen Huang supports US development of open models, while Anthropic CEO Dario Amodei has argued for greater caution. Their positions are relevant to the debate but come from executives whose companies have different commercial interests.

For governments, the choice is not simply whether to permit or prohibit open models. The report describes a range of possible approaches between a blanket ban and unrestricted release. Each involves trade-offs between limiting potential misuse and distributing technical capability beyond a few firms or states.

From Closed Services to Open Weights

Closed systems such as those accessed through ChatGPT or Claude are generally used through a company’s website, app or API. Users can submit prompts and receive outputs, but usually cannot examine or alter the underlying model. With an open-weight release, users can obtain the model’s parameters and run or adapt them, subject to their own computing resources and technical skill.

MIT Technology Review frames this as a continuing argument, not a settled policy question. David Gray Widder, an assistant professor at the University of Texas at Austin School of Information, told the publication that views on openness are often shaped by broader views about AI. People emphasizing safety focus on misuse; people emphasizing practical benefits argue that powerful technology should not remain under the control of a few large companies.

The article also describes an international dimension. It says sophisticated hacking models are currently made in the United States, raising questions about who can access tools that might support cyber defenses. The possibility that open models could alter that balance is presented as a scenario, not a demonstrated outcome.

““I think a lot of the way you come down on the openness debate is influenced by how you come down on the AI debate.””

— David Gray Widder, assistant professor at the University of Texas at Austin School of Information

How Much Risk Can Be Measured

The available material does not establish whether GLM-5.3 has enabled a cyberattack, how likely such use is, or how its risks compare with those of closed models. Anthropic’s report raises a concern; the article does not provide an independent evaluation that resolves it. Nor does it specify which safeguards, if any, would reliably limit misuse after model weights have been downloaded.

It also remains uncertain how quickly open models will close the reported capability gap, and whether the leaderboard comparison captures performance across all tasks. The costs and security benefits of running models locally will vary by organization, hardware and use case. The article lays out competing arguments rather than offering a single measured calculation of whether benefits outweigh risks.

Policy Choices After the Warning

The immediate next step is closer scrutiny of the evidence behind risk claims and of the practical benefits of open-weight systems. Policymakers weighing restrictions will have to decide which releases, capabilities or uses warrant limits, and whether any rules can account for models that users may modify after downloading. The report does not announce a government decision or a specific regulatory timetable.

Further model evaluations, disclosures from developers and independent research could help clarify the risks and the effectiveness of safeguards. For now, the central question remains how to preserve access for research and useful applications while reducing opportunities for harmful use—a balance the report says cannot be resolved by treating all open models as either safe or dangerous.

Key Questions

What is an open-weight AI model?

It is a model whose weights or parameters can be downloaded. Users may run it on their own hardware and, with the necessary expertise and resources, adapt it. The term does not necessarily mean that its training data or full development process is public.

What risk did Anthropic raise about GLM-5.3?

Anthropic argued that the model poses a cyberattack risk because it is powerful and freely available to download and modify. The report does not establish that GLM-5.3 has been used in an attack.

Why do researchers and businesses use open models?

They can support experiments into a model’s internal behavior, allow fine-tuning and model combinations, and be run locally. Local deployment may help some businesses keep sensitive information on their own systems; costs and security outcomes depend on the use case.

Does the report conclude that open models’ benefits outweigh their risks?

No. The analysis describes competing benefits and risks but does not offer a single verdict. It says policy choices involve trade-offs between safety and broader access.

Source: rss