AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Open models have real benefits. Do they outweigh the risks?

Collected Oct 13, 2026

Late last month, Anthropic released a report arguing that the open-weight model GLM-5.3, released in August by the Chinese company Z.ai, poses a major cyberattack risk. The report's core concern is that the model is freely downloadable and modifiable by anyone. The assertion lands in the middle of a long-running argument over whether frontier AI models should be released openly or held behind the doors of the frontier labs, and that argument now plays out largely in public, with substantial commercial stakes attached.

Nvidia chief executive Jensen Huang advocates enthusiastically for the United States to develop more open models; Anthropic's Dario Amodei argues for a more cautious approach. Both have obvious commercial positions in the outcome. David Gray Widder, an assistant professor at the University of Texas at Austin School of Information, frames the disagreement as downstream of a larger one: where you land on openness tends to track where you land on AI itself. Safety-focused voices warn that open models are easy for bad actors to misuse. Those more focused on AI's practical potential argue the capability should not sit behind the closed doors of a few large companies.

The practical difference between the two camps comes down to distribution. A closed model such as those behind ChatGPT or Claude is reachable only through a website, app or API. An open model can be downloaded, with its internal structure encoded in a long series of numbers called the model weights, made publicly available. The term open-weight is often used to distinguish this from models that also disclose more, such as training procedures. As of publication, closed models remain more capable: the top eight entries on the Artificial Analysis LLM leaderboard are all closed, though the article notes open models are only a handful of months behind.

Benefits cited for open models are concrete. Because they are downloadable, they can be run on local machines or drives, and retrained to behave differently given the expertise and compute. Academic researchers can inspect internal workings and the chain of thought a model uses before answering, combine multiple open models into new systems, compress them for efficiency, and fine-tune them. A recent Mozilla report found open models typically cost less than closed ones, even when an external provider runs them, and running a model locally is often the best way to handle sensitive or proprietary data.

The broader argument is about power. Elizabeth Seger, senior policy advisor at the Tony Blair Institute for Global Change, calls preventing mass concentration of power the clearest and biggest argument for open models, framing the question of who controls AI infrastructure as comparable to who controls information on the internet. In one possible future she sketches, frontier systems are controlled entirely by select people, companies or governments whose influence over the world grows with the systems. Today's most sophisticated hacking models originate in the United States, so Washington and US companies effectively decide whether other countries can use them for cyber defense. Hamish Hobbs, director of AI policy at the Centre for Long-Term Resilience, argues a world of individual, loyal, secure AI systems would be less concerning than one where people must trust an inscrutable government system.

Open models are not automatically a geopolitical fix. Paul Azunre, cofounder of Ghana NLP, which builds open-source tools for low-resource Ghanaian languages, describes a pattern in which companies release models, watch how the community modifies them, then fold those improvements into their own products without credit or compensation. Mark Zuckerberg has himself said that integrating community improvements into Meta's stack is a key benefit of releasing open models. Azunre calls the ideal of open source noble but says it tends to be implemented as an extractive approach.

On risk, the article is blunt that closed models are hardly safe. It cites an autonomous hack OpenAI agents executed against Hugging Face in July, and an Anthropic risk report documenting extensive hacking operations carried out with Claude's assistance. Anthropic screens queries to its closed models for hacking, biological weapons and other risks, and can shut down accounts when users evade those filters. A downloaded model has no such filter and no way to revoke access. Models are also trained to refuse harmful requests, a defense intrinsic to the weights, but researchers have shown it is surprisingly easy to retrain a safe open model into compliance with the behavior it was taught to avoid.

Some therefore argue open models are simply too dangerous to release, imagining a model capable enough to guide someone through synthesizing a virulent pathogen. But the article cautions against equating open with dangerous and closed with safe. Closed-model defenses are imperfect and misuse is common. Bad actors may also prefer closed models, which are more accessible than comparably capable open ones that require expensive chips to run locally and more still to retrain.

Experts who spoke to MIT Technology Review drew an analogy to open-source software. Linux and similar tools are often considered safer than closed counterparts because a large community probes and patches vulnerabilities. Communities can test open models for flaws more easily, though fixing them is trickier than editing a line of code. Open models also spread cyber defense more widely, since hacking tools and defensive testing overlap. Anthropic's Project Glasswing lets trusted companies use its most capable models for defense, but many firms are left out. After the Hugging Face hack, the company noted it had to use an open model to analyze the incident because questions about it tripped closed models' security filters.

Still, many experts see open models as riskier. Seger points to low-probability, catastrophic incidents that would be disastrous if they occurred, such as a highly capable open hacking model released before defenses could be hardened. Because open models have few guardrails and are decentralized, misbehavior by the models themselves could be hard to detect.

Attention is shifting toward middle ground. Filters and monitoring do not transfer to downloaded models, so researchers are exploring training runs that never expose a model to data teaching bad behavior, though purging datasets without hurting performance remains hard. Others pursue tamper-resistant models: repeatedly retraining a model to be unsafe while reinforcing its resistance, in theory producing lasting fortification. A recent preprint found no current approach withstands all tampering attacks, though lead author Saad Hossain, an engineer at Deep Breathe, is optimistic given how active the research area is.

Openness itself has limits. Hobbs says deconcentration benefits are genuine but insufficient on their own. Open weights are a bare minimum; training data and procedure details would teach researchers far more, and companies rarely disclose them. Hardware matters too, since weights are useless without chips capable of running them, and those chips have grown prohibitively expensive as AI companies buy them up. Widder calls openness a floor, not a ceiling. Fuller openness, including accessible hardware, would carry more risk, since anyone could then be a hacker. Staged release, where a model starts closed and becomes more freely accessible as safety is assured, is a widely floated mitigation. Implementing it requires balancing the risks and benefits carefully.

Why it matters: For teams choosing a model stack, this is a live trade-off rather than a settled question. Open weights can lower cost, enable fine-tuning, local processing of sensitive data, and inspection of a model's reasoning, while closed models currently lead capability leaderboards and carry vendor-side filters and account controls that downloadable weights do not. If regulators or labs move toward staged release or tamper-resistance requirements, expect release timing, licensing and disclosure terms to become procurement criteria. That is inference, but it follows from the direction of the mitigation research and the policy proposals described.

Read at MIT Technology Review · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

MIT Technology Review Explains: Let our writers untangle the complex, messy world of science and technology to help you understand what’s coming next. You can read more from the series here. Late last month, Anthropic released a report arguing that the model GLM-5.3 poses a major cyberattack risk. The report’s core concern is that the…