The BaseNewsMIT Technology Review

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

MIT Technology Review's explainer weighs open-weight AI models, prompted by Anthropic's report arguing that the open model GLM-5.3 poses a major cyberattack risk. It sets benefits (research access, lower costs, less concentrated power) against risks (no filters or account shutdowns for downloaded models, easily removed safety training), and reports experts' middle paths such as staged release.

Open the source ↗
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

Anyone can download an open model and run it themselves, which can cut costs and keep data in-house, but no company can filter prompts or cut off a misuser.

For lawmakers

Between banning open models and releasing them freely, the article describes staged release, where models start closed and open up as safety is assured, an approach many experts back.

For researchers

Researchers need open models to inspect inner workings, yet retraining can strip safety training, and a cited preprint finds no current tamper-resistance method withstands all attacks.

The question it raises

Who should control the infrastructure AI runs on: a few labs and governments, or everyone? The experts quoted treat openness as a starting point, not a full answer.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“But it would be a vast oversimplification to deem open models dangerous and closed models safe.”
MIT Technology ReviewThe explainer's central argument, in the outlet's own words, after setting out the case that open models are riskier than closed ones.
Found word for word in the source
“But when someone downloads a model and runs it on their own hardware, there’s no filter checking whether their prompts are safe, and there’s no way to revoke their access to the model.”
MIT Technology ReviewIn the outlet's own words, on why query filters and account shutdowns, used for closed models, do not carry over to downloaded models.
Found word for word in the source
“I think we’re at this juncture where the question of who controls the infrastructure that allows AI to operate is as big of a question as who gets to control information on the internet.”
Elizabeth Seger, Senior policy advisor, Tony Blair Institute for Global ChangeElizabeth Seger of the Tony Blair Institute for Global Change, as quoted, on why control of AI infrastructure is at stake in the open-versus-closed debate.
Found word for word in the sourceAs quoted in the text we read, not the speaker’s own text or a transcript of their words
“But the way it’s actually implemented in practice, it tends to be a very extractive approach.”
Paul Azunre, Cofounder, Ghana NLPPaul Azunre of Ghana NLP, as quoted, after calling the ideal of open source noble; the article reports Azunre's view that AI companies adopt community improvements without credit or pay.
Found word for word in the sourceAs quoted in the text we read, not the speaker’s own text or a transcript of their words
“Openness is a floor, not a ceiling, for what we need to do to ensure that our technological futures benefit everyone.”
David Gray Widder, Assistant professor, University of Texas at Austin School of InformationDavid Gray Widder of the University of Texas at Austin, as quoted, on what openness does and does not achieve.
Found word for word in the source; punctuation differsAs quoted in the text we read, not the speaker’s own text or a transcript of their words

The record

Questions
Are AI companies putting stronger safeguards into practice? · How is AI used in attacks, defence and war? · Who can demand evidence, delay a release or stop an unsafe deployment?
People
Grace Huckins (Wrote it) · David Gray Widder (Speaks, Assistant professor, University of Texas at Austin School of Information) · Elizabeth Seger (Speaks, Senior policy advisor, Tony Blair Institute for Global Change) · Hamish Hobbs (Speaks, Director of AI policy, Centre for Long-Term Resilience) · Paul Azunre (Speaks, Cofounder, Ghana NLP) · Saad Hossain (Speaks, Engineer, Deep Breathe; lead author of a preprint on tamper-resistant models) · Jensen Huang (Mentioned, CEO, Nvidia) · Dario Amodei (Mentioned) · Mark Zuckerberg (Mentioned)
What we read
We read the whole source.
Found by
The Radar, run of October 14, 2026, via mit-technology-review-ai (tier 2: a major outlet’s own reporting)
Cite this
MIT Technology Review, “Open models have real benefits. Do they outweigh the risks?”, October 13, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/mit-technology-review-2026-10-13-open-models-benefits-risks
Our notesAbout Anthropic or Claude: the piece starts from an Anthropic report on GLM-5.3 and cites a recent Anthropic risk report on hacking operations that used Claude. Both are described second-hand, the article does not say whether they are one document, and we have not read either. The same goes for its accounts of the OpenAI and Hugging Face incident, the Mozilla cost report, the leaderboard ranking and the tamper-resistance preprint. The article also says what the chief executives of Nvidia and Anthropic stand to gain commercially; those are the author's characterizations, and we do not repeat them. The experts' titles are as the article gives them, and their quotations are as the outlet reports them.