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

Is open-source AI safer or riskier? I'll map the trade-offs between open and closed models.

Neither is categorically safer — open-source and closed/proprietary AI have different safety trade-offs, and research suggests the answer depends on which risk you mean.

1. Why open-source is argued to be safer

•Transparency and auditability: Proponents argue openness democratizes capabilities, enables independent safety auditing, prevents monopolistic concentration of power, and accelerates research. The open-source community contends that "security through obscurity" is a fallacy and that only a global community can identify and fix systemic risks.

•More eyes: Meta's argument for Llama 3.1 was that open-source AI is safer due to transparency and community review. A common framing is: open camp believes transparency: more eyes make it safer vs closed camp believes control: fewer copies make it safer.

•Faster patching and compliance: Open-source models also encourage a rapid patching of vulnerabilities. With access to weights and training data, models are inspectable and compliant by default compared to black-box APIs — though some 2026 claims that 87% of privacy incidents came from closed models are not widely substantiated.

•Prevents single point of failure: Openness eliminates the single point of failure and prevents a small group of engineers from having unilateral control.

Some surveys of the research literature conclude that, overall, the benefits of open-source Gen AI outweigh its risks when developed responsibly.

2. Why open-source is argued to be riskier

•Perceived risk is higher: In a McKinsey / Mozilla Foundation survey of 700+ tech leaders, respondents consider open source AI tools to be riskier than proprietary AI for most types of AI risks — cybersecurity (62%), regulatory compliance (54%), and intellectual property (50%) were top concerns. Organizations cite concerns about cybersecurity, regulatory compliance and intellectual property when engaging with AI tools.

•Irreversible release and guardrail removal: Once weights are public, they cannot be un-released, and fine-tuning can remove safety guardrails. Adversaries may directly manipulate a model's weights, or modify or remove parts of a model to create a jailbroken or "uncensored" variant. This is applicable to open-source models. This leads to a 'safety gap' — the difference between a model with intact safeguards and one stripped of them — which widens as model scale increases.

•No oversight after download: Politicians and security officials warn that powerful openly accessible models fundamentally alter the security situation because they lower the barrier for cybercriminals and state actors to automate attacks or run influence campaigns. Criminals could operate open models locally, modify them, and use them without the oversight of the original provider.

•Security risks remain: Risks include exposure to backdoor attacks, data poisoning, and unauthorized access. Technically, while open-source AI may provide a higher degree of transparency, other risks including hallucinations, capacity to cause harm, IP infringement and privacy risks are not necessarily mitigated by openness.

Empirical tests are mixed: One audit found DeepSeek-R1, an open-source reasoning model, failed 58% of adversarial safety tests, far worse than its closed peers, while another evaluation by Telus Digital found open source models are not inherently less safe than closed ones, with Zhipu AI's GLM 4.7 outperforming many proprietary alternatives.

3. Why closed/proprietary is argued to be safer

•Centralized control: Closed models allow for centralized control over safety filters and alignment. Companies argue limiting access prevents malicious actors from exploiting capabilities for disinformation, cyberattacks or autonomous weapons, and allows more rigorous safety testing before deployment.

•Managed safety: The main advantage of closed models is that the provider handles performance, safety, and reliability for you, often with mature safety systems that lead on the hardest reasoning tasks. They integrate safeguards to comply with regulations, such as content moderation filters and bias detection.

Recent analysis notes several studies demonstrate that closed models typically demonstrate fewer safety and security risks compared to open source, though closed models are also vulnerable to jailbreaking techniques.

4. Why closed is argued to be riskier

•Opacity and vendor lock-in: When a business integrates closed-source AI it cedes control of its operational intelligence to a third party, creating precarious dependency. Risks include dependency on closed, opaque providers with limited portability and vendor lock-in.

•Hidden risks: Without inspectability, vulnerabilities and biases can remain hidden and limit accountability, whereas open weights encourage shared responsibility.

Deploying an open source solution is not inherently any riskier than deploying a proprietary one if proper vetting is done — the same vendor evaluation criteria apply.

Bottom line: Safety ≠ Openness

A widely cited synthesis: Neither label guarantees safety. Open systems can let outside researchers inspect, test and modify models in ways closed systems may not permit, but the same accessibility can raise concerns about misuse or removal of safeguards. A model does not become safe simply by being open, nor does a closed model become safe simply by keeping its weights private.

Many experts conclude the ecosystem needs both, and that hybrid models that combine transparency in research with controlled deployment may offer the most pragmatic path forward, with trade-offs framed as:

•Innovation vs misuse: Open speeds research but expands attack capability

•Transparency vs control: Open can be audited, closed can be throttled

•Competition vs concentration: Open prevents monopoly, proprietary can fund large safety teams

2.[Open source](https://www.mckinsey.com.br/~/media/mckinsey/business functions/quantumblack/our insights/open source technology in the age of ai/open-source-technology-in-the-age-of-ai_final.pdf)

Citations (4)

1arxiv.orgAcademic
Risks and Opportunities of Open-Source Generative AI
Backs: “Prevents single point of failure: Openness eliminates the single point of failure and prevents a small group of engineers from having unilateral control. Some surveys of the research literature conclude that, overall, the benefits of open-source Gen AI outweigh its risks when developed responsibly.”
2mckinsey.com.brCompanies
Open source
Backs: “Perceived risk is higher: In a McKinsey / Mozilla Foundation survey of 700+ tech leaders, respondents consider open source AI tools to be riskier than proprietary AI for most types of AI risks — cybersecurity (62%), regulatory compliance (54%), and intellectual property (50%) were top concerns. Organizations cite…”
3osna.fmNot yet classified
Politicians Warn Open AI Models Pose Critical Security Risks to National Safety
Backs: “This leads to a 'safety gap' — the difference between a model with intact safeguards and one stripped of them — which widens as model scale increases. No oversight after download: Politicians and security officials warn that powerful openly accessible models fundamentally alter the security situation because they…”
4lambham.comCompanies
Open vs. Closed AI Models: Navigating the Epoch Debate
Backs: “Centralized control: Closed models allow for centralized control over safety filters and alignment. Companies argue limiting access prevents malicious actors from exploiting capabilities for disinformation, cyberattacks or autonomous weapons, and allows more rigorous safety testing before deployment.”