I'll break down the safety tradeoffs between open-source and closed AI.
Neither is categorically safer — they create different risk structures, and safety depends on capability level, deployment context, and governance.
1. Why open-source is argued to be safer
•Transparency and auditability: Open models can be audited, making it easier to identify biases or errors. This contrasts with closed systems, where lack of visibility can obscure risks and limit accountability. Proponents argue open-source AI is safer due to transparency and community review.
•Security through community: The open-source community argues that "security through obscurity" is a fallacy, and that global researchers can identify and fix systemic risks. From this perspective, open-source eliminates a single point of failure and prevents unilateral control by a small group of engineers.
•Data sovereignty and resilience: Open-weight AI can provide data sovereignty, operational independence, security-research freedom, customization, deeper auditability and resilience when an external service is unavailable. For example, a hospital investigating ransomware or a bank examining malware may not be able to send sensitive evidence to an external AI provider, illustrated when Hugging Face needed to run a model locally to analyze attack artifacts because commercial APIs blocked the requests for safety reasons.
•Avoiding vendor lock-in and black-box risk: Closed-source proprietary models introduce opaque "black-box" architectures, external data privacy risks and dependency on closed, opaque providers with limited portability and risk of pricing changes or API deprecations. Closed models can't be independently audited for bias, backdoors, or capability regressions.
Proponents also note open models, with inspectable weights and training data, allow third-party verification in sensitive domains like justice or healthcare.
2. Why open-source is seen as riskier
Survey data shows leaders perceive open source as riskier for most risk types:
> Respondents consider open source AI tools to be riskier than proprietary AI for most types of AI risks. The most relevant AI risks cited include cybersecurity (62 percent of respondents), regulatory compliance (54 percent), and intellectual property (50 percent)
The same figures appear in McKinsey's 2025 survey of 700+ technology leaders: concerns about cybersecurity (62 percent), regulatory compliance (54 percent), and intellectual property (50 percent).
Specific risk mechanisms:
•Irreversibility: Once weights are downloadable, a user can potentially operate the model locally, fine-tune it, alter safeguards, and continue operating it even if the developer later decides the model is unsafe. Publishing powerful model weights is therefore not just software distribution. For sufficiently capable systems, it may be an irreversible transfer of capability. New safeguards cannot automatically be applied to privately hosted versions.
•Misuse at scale: Making models open accelerates diffusion, but also increases the potential for misuse. Some open-source models have been used to generate misleading content at scale in automated disinformation campaigns. Researchers warn hackers can easily commandeer computers operating open-source LLMs outside the guardrails of major platforms.
•Cyber and biosecurity capability: A Chinese threat actor combined DeepSeek, an open AI model, with an autonomous agent tool to attempt intrusions into more than 460 servers worldwide. In an August 2026 evaluation of the open-weight GLM-5.2, the model refused none of the offensive security or biological tasks, while the organization emphasized a self-hoster can remove whatever safeguards the developer originally included.
•Operational burden: Open models offer greater transparency, but it still requires rigorous security management. When a model is deployed locally, the organization becomes responsible for access management, encryption and monitoring, and poor configuration can expose sensitive information.
3. Why closed/proprietary is not automatically safe
•Closed models can still fail dangerously: Closed models can still behave unexpectedly, find vulnerabilities, circumvent containment, misuse tools, gain excessive autonomy or take actions their developers did not intend.
•Concentration risk: A closed model can create concentrated risk inside a small number of extremely powerful companies. An open model can distribute powerful capability across an enormous number of actors. Those are different risk structures. Neither should be ignored.
•Opacity hampers regulation: The primary regulatory challenge with proprietary AI lies in its "black box" nature without external visibility into decision-making, training data, or logic, making it difficult for regulators to assess compliance or verify safety claims.
4. The emerging consensus: Risk-tiered, hybrid approach
Current debate has moved from "open vs. closed" to what should be open, and when:
•Tier 1 - Broad Open Release: Models below serious misuse thresholds freely downloadable
•Tier 2 - Controlled Weight Access: More powerful models available to vetted researchers/companies
•Tier 3 - Hosted Access: Models crossing cyber/biological/autonomy thresholds remain accessible only through monitored infrastructure
•Tier 4 - Frontier Restricted Systems: Extremely capable systems require independent assessment and potentially government oversight
Under this view, an open model should not receive less safety evaluation because its developer cannot control it after release — arguably it should receive more.
In practice, most enterprises are moving toward a hybrid stack. Three-quarters of respondents expect to increase open source use, but more than 70% say they are open to either open or proprietary technologies depending on the layer of the AI tech stack.
> Bottom line: For routine, low-risk tasks, transparency, cost, and data-control advantages often make well-managed open models safer and more trustworthy. For frontier-level capabilities (advanced cyber offense, bioweapon-relevant biology, autonomous agents), the irreversibility of open weights makes open release riskier than a controlled, monitored API, even though closed systems carry their own opacity and concentration risks.
3.[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)