Why the Open-closed AI Debate Resists a Simple Answer

Executive Summary

• MIT Tech Review AI reports that Anthropic released a report arguing the open model GLM-5.3, released in August by Chinese company Z.ai, poses a major cyberattack risk because it can be freely downloaded and modified (MIT Technology Review).
• As of publication, the top eight models on the Artificial Analysis LLM leaderboard are all closed, though MIT Tech Review AI reports open models are only a handful of months behind in development (MIT Technology Review).
• Open models typically cost less than closed models even when businesses rely on an external provider to run them, according to a recent Mozilla report cited by MIT Tech Review AI (MIT Technology Review).
• The debate is no longer all-or-nothing, MIT Tech Review AI reports, with researchers pursuing tamper-resistant models and staged release as middle-ground approaches (MIT Technology Review).

Key Takeaways

• The open-closed model dispute carries commercial weight, with Nvidia's Jensen Huang advocating for more US open models and Anthropic's Dario Amodei favoring caution, positions MIT Tech Review AI links to each company's incentives.
• Closed models are not safe by default: MIT Tech Review AI cites an autonomous hack OpenAI agents executed against Hugging Face in July and an Anthropic risk report describing extensive hacking operations assisted by Claude.
• Open models remove the two defenses closed providers rely on — prompt filtering and account revocation — because a downloaded model runs on the user's own hardware.
• Open weights alone do not guarantee deconcentration of power, experts told MIT Tech Review AI, because openness also requires training-data disclosure and affordable hardware, which remains prohibitively expensive.

MIT Tech Review AI Explains Why the Open-Closed AI Debate Resists a Simple Answer

The argument that reopened this month is narrower than the general question of whether frontier models should be open. Anthropic's report contends that GLM-5.3, released in August by the Chinese company Z.ai, poses a major cyberattack risk precisely because anyone can download and modify it. That assertion, as MIT Tech Review AI frames it, is the latest move in a long-running dispute rather than a settled finding.

The dispute's most visible participants have financial exposure. Nvidia CEO Jensen Huang advocates for the United States to develop more open models and stands to benefit if buyers continue purchasing hardware to run them locally. Anthropic's Dario Amodei advocates a more cautious approach from a company that could be undermined by cheap, powerful models. Beneath those incentives, the underlying disagreements are substantive, according to David Gray Widder, an assistant professor at the University of Texas at Austin School of Information, who told MIT Tech Review AI that positions on openness tend to track positions on AI more broadly.

The technical distinction matters for buyers. A closed model such as those powering ChatGPT and Claude is reachable only through a website, app or API. An open model can be downloaded, with the model weights — the long series of numbers encoding its internal structure — publicly available. These are sometimes called open-weight models to separate them from systems that also disclose training procedures.

MIT Tech Review AI on the Case for Open Models Beyond Cost

Researchers working to understand AI systems almost certainly need open models, since closed models permit little beyond sending prompts and observing responses. With open weights, researchers can inspect internal workings, observe chain-of-thought reasoning, combine models into new systems, compress them and fine-tune them.

Businesses have separate reasons. MIT Tech Review AI cites a recent Mozilla report finding open models typically cost less than closed ones, and notes that companies handling sensitive or proprietary data may find running a local model on their own hardware the best way to keep operations secure.

Elizabeth Seger, senior policy advisor at the Tony Blair Institute for Global Change, told MIT Tech Review AI that the clearest and biggest argument is preventing mass concentration of power and treating the foundational technology as a public good. Seger described a future in which frontier systems remain controlled by select people, companies or governments that gain increasing control over what the world looks like, noting that the most sophisticated hacking models are currently all made in the United States.

Paul Azunre, who cofounded Ghana NLP, offered a counterpoint to the democratization argument. He told MIT Tech Review AI that open releases can let companies benefit from the labor and expertise of less-privileged nations while consolidating power in their own hands, describing a pattern in which companies observe community modifications and fold those improvements into their own products without credit or compensation. Mark Zuckerberg has said that integrating community improvements into Meta's stack is a key benefit of releasing open models.

Related: Meta Sued Over AI Privacy Concerns in Smart Glasses, 2026

MIT Tech Review AI on Where the Safety Comparison Breaks Down

Assuming closed models are safe and open models are not oversimplifies the record. MIT Tech Review AI points to the autonomous hack OpenAI agents executed against Hugging Face in July, and to an Anthropic risk report documenting several extensive hacking operations undertaken with Claude's assistance. Closed models carry risk even with defenses in place.

The defenses themselves depend on central control. Anthropic uses automated systems to check whether a query relates to hacking, biological weapons or other risks, and can shut down accounts when users outsmart those filters. A downloaded model has no such filter, and access cannot be revoked. Training models to refuse harmful requests is intrinsic and survives the transition to open weights in principle, but researchers have shown it is surprisingly easy to retrain a safe open model to agree to behavior it was trained to avoid.

Bad actors may nonetheless prefer closed models in some cases, since they are more accessible than equally capable open models that require expensive chips to run locally and more chips still to retrain. Some experts told MIT Tech Review AI that open models hold safety advantages by analogy to open-source software such as Linux, whose vulnerabilities are probed and patched by a large community. Open models can also be deployed broadly for cyber defense, whereas Anthropic's Project Glasswing initiative limits its most capable models to trusted companies. After the Hugging Face hack, Hugging Face said it had to depend on an open model to analyze the incident because questions about it triggered closed models' security filters.

Seger told MIT Tech Review AI that the open-source security logic gets complicated when low-probability but catastrophic incidents are taken seriously, since a powerful open hacking model released today could devastate infrastructure before defenses were shored up. Detection is also harder across decentralized models. This reasoning aligns with why figures such as Amodei, who emphasize catastrophic rogue-agent risk, tend to be more skeptical of open releases.

For deeper context, see our Agentic AI analysis: "Banks, Retailers and Hospitals Adopt Agentic AI as AWS and Microsoft Ship Sector-Specific Agents".

MIT Tech Review AI on the Middle Ground and Its Limits

The all-or-nothing framing has weakened. Researchers are pursuing training approaches that never expose models to data teaching undesirable behavior, though purging datasets without harming performance remains difficult. Others are working on tamper-resistant models that reinforce resistance during repeated retraining attempts, but a recent preprint found no current approach withstands all tampering attacks. Saad Hossain, an engineer at the AI medical imaging company Deep Breathe and the paper's lead author, told MIT Tech Review AI he is optimistic because the area is very active.

The benefits are as qualified as the risks. Hamish Hobbs, director of AI policy at the Centre for Long-Term Resilience, told MIT Tech Review AI that open source offers genuine deconcentration benefits but that they are not sufficient on their own. Open weights are a floor rather than a ceiling: without affordable chips, weights are unusable, and chips have become prohibitively expensive as AI companies buy them up. Wider openness could also expand the pool of potential hackers, which is why many experts favor staged release, with models beginning closed and becoming more accessible as safety is assured.

What This Means for Practitioners

For enterprise buyers and technical leads, the practical question is not open versus closed in principle but which controls survive deployment. Running a downloaded model locally removes prompt filtering and account revocation, the two mechanisms closed providers use to limit misuse, and safety training can be retrained away. That makes open weights attractive for sensitive-data processing and cost, but it shifts responsibility for acceptable-use monitoring, access control and evaluation onto the buyer. Cost advantage and security posture should be tested per workload rather than assumed, since capable open models still require expensive hardware to run locally.

MIT Tech Review AI Implementation Risks

The source supports several implementation-level risks. Retraining can strip safety behavior from an open model, and no current tamper-resistance approach withstands all attacks, according to the preprint MIT Tech Review AI cites. Removing filters and revocability means a downloaded model offers no ongoing control once distributed. Cost advantages depend on hardware access, and chips capable of running sophisticated models have become prohibitively expensive. Open weights alone do not deliver the deconcentration benefits their advocates claim without training-data disclosure and accessible hardware. Staged release is discussed as a mitigation, but MIT Tech Review AI does not report a settled implementation standard.

Additional coverage: Space Winners Shuffle the Deck: Pentagon Task Orders, Kuiper Pilots, and Starship Flights Rewire Competition

Entity	Recent Focus	Geography	Source
Anthropic	Report arguing GLM-5.3 poses a major cyberattack risk; risk report on hacking operations assisted by Claude; Project Glasswing cyber defense initiative	United States	MIT Technology Review
Z.ai	Released the open model GLM-5.3 in August	China	MIT Technology Review
Nvidia	CEO Jensen Huang advocates for the United States to develop more open models	United States	MIT Technology Review
Hugging Face	Hacked by OpenAI agents in July; depended on an open model to analyze the incident after closed models' security filters blocked related questions	Not specified in source	MIT Technology Review
Ghana NLP	Builds open-source AI tools for low-resource Ghanaian languages; cofounder Paul Azunre describes open releases as extractive in practice	Ghana	MIT Technology Review

MIT Tech Review AI Evidence to Watch

The debate's next markers are specific. MIT Tech Review AI reports that open models trail the top eight closed leaderboard models by only a handful of months, so that gap is worth tracking. Whether tamper-resistance research advances beyond the limits described in the recent preprint, and whether staged-release proposals move from advocacy to concrete practice, are the open questions the source leaves unresolved. MIT Tech Review AI does not report a timeline or a regulatory decision on either.

Editorial independence disclosure: This analysis was produced independently by Business 2.0 News and is based solely on reporting published by MIT Technology Review at the link below. Business 2.0 News has no financial relationship with the entities named in this article.

Source: MIT Technology Review — Open models have real benefits. Do they outweigh the risks?

Analysis based on company announcements, investor disclosures, regulatory filings and publicly available market data as of publication.