Is open-source AI safer or riskier than closed/proprietary AI?

This is a genuinely contested question among AI researchers, policymakers, and industry leaders, with reasonable arguments on both sides. Let me search for the current state of this debate to give you a well-grounded answer.## There's No Consensus—It Depends on What Risk You're Measuring

This is one of the most actively contested debates in AI, with credible experts and evidence on both sides. As of 2026, it's escalated into a high-profile industry split rather than being settled.

The Case That Open-Source AI Is Riskier

1. Irreversibility of harm. Once weights are released, there's no way to pull them back. Anthropic argues that open weight releases mean "the same openness underpinning these benefits precludes many of the safety measures that closed model developers can use to detect and disrupt misuse, iterate on safeguards as vulnerabilities emerge, control user access and withdraw models," and once released "these options are lost permanently: safeguards can be removed, and copies can be downloaded, redistributed, and run on private systems beyond monitoring."

2. Safeguards are trivially removable. A major 2026 study found that all 21 leading open-weight AI models tested could be modified to bypass safety protections, with an international team led by the University of Waterloo and FAR.AI finding they could all be tampered with despite built-in safeguards.

3. Concrete biosecurity and abuse evidence. RAND research found that open-weight LLMs can be modified by users, creating benefits but also biosecurity risks, and removing an LLM's safety training is highly feasible. Separately, the International AI Safety Report notes that modified open-weight image-generation models have become the most common tools used for creating AI-generated child sexual abuse material.

4. Governance gaps. Open models shift risk to a decentralized, hard-to-police landscape: AI has rapidly transitioned from a centralized, platform-controlled service to a decentralized, user-run infrastructure, with tens of thousands of publicly accessible computers now running open-source LLMs outside the security controls imposed by major AI platforms. This is compounded by disinformation generated at scale via exposed open-source LLMs, often from residential IPs that make malicious traffic hard to distinguish from ordinary activity.

The Case That Open-Source AI Is Safer (or at Least Not Worse)

1. Closed models aren't actually secure either. A widely cited academic analysis pushes back on the assumption that closed is inherently safer: closed models are also vulnerable to jailbreaking, and their safeguards are easily bypassed through fine-tuning via accessible services, prompting tricks, ciphers, or foreign-language instructions—completely preventing undesirable behavior might not even be possible. The same source concludes it is not clear that closed models are definitively "safer" than open-source models.

2. "Restrictable access" is a myth. The assumption that closed models can always have access revoked ignores that closed models can be compromised via hacking, leaks, reverse engineering, or duplication—and on-premise deployments may not be retractable at all, as when Mistral's 70B parameter model was leaked by an early customer.

3. Concentration-of-power risk. Meta's Mark Zuckerberg has made this the center of his 2026 argument: broadly available models are structurally safer because they deny any single actor a monopoly on capability, while Altman's camp argues broad availability is dangerous because it denies any safety body a meaningful gate. Nvidia's Jensen Huang has echoed this, and a real-world incident lent it weight: after OpenAI's agents reportedly breached Hugging Face, Hugging Face said it was forced to use an open-weight model developed in China to diagnose and defend from the hack because of the closed models' internal guardrails.

4. Transparency enables scrutiny. Wider openness generally supports oversight: in general, the more open a model is—including access to training data, code, documentation, and compute infrastructure—the greater the benefit for innovation and safety oversight. Similarly, releasing weights openly allows far more actors to perform deeper technical analysis to spot flaws and biases, even though it also risks propagating those same flaws downstream.

Where the Evidence Actually Lands

The most balanced technical assessments (including the 2026 International AI Safety Report, produced by a broad international expert panel) land in the middle rather than declaring a clear winner:

Bottom Line

There is no single correct answer—it depends on which risk dimension you weight most heavily:

Even the researchers making the most systematic case tend to avoid declaring one paradigm categorically "safer," instead noting that some models might enable acute misuse if deployed with open weights, others might concentrate power or hinder open science if deployed closed, and others still might pose major risks regardless of deployment type—meaning the safety question is less "open vs. closed" and more "how capable and dangerous is this specific model, and what safeguards actually work for it."