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

Whether open-source AI is safer is a live debate — I'll weigh the evidence for both sides.

It is not a simple either/or - research and regulators describe the evidence as genuinely mixed. Both open-source and closed/proprietary AI have safety advantages and safety risks, and "safer" depends on what threat you are worried about.

1. Arguments that Open-Source is Safer

Transparency vs Security Through Obscurity

Open-source advocates argue "security through obscurity" is a fallacy. By keeping models closed, the industry is ignoring systemic risks that only a global community of researchers can identify and fix.

From this perspective, open-source AI is actually safer because it eliminates the single point of failure and prevents a small group of engineers from having unilateral control over the world's most powerful cognitive tools.

Specific benefits cited:

•Independent auditing: Transparency allows for independent scrutiny and auditing of AI systems’ behaviors and ethics — and when we leverage the existing interest and drive of the masses, they will find the problems and mistakes. In a closed system, the "safety" of a model is determined by a private board; in an open system, the global research community can audit the model for biases, vulnerabilities, and hallucinations.

•Fostering innovation and reproducibility: Benefits of open model weights include fostering innovation, enhancing transparency and supporting reproducibility in AI research.

•Democratization and sovereignty: Open-source democratizes the technology, allows quantization to run on consumer hardware and deep fine-tuning on private data without that data leaving a secure environment, and prevents "digital colonialism" where values of a foreign corporation are baked into another country's systems.

•Proponents argue benefits outweigh risks: Some papers argue open source reduces the likelihood of existential risk by contributing to technical alignment and maintaining balance of power, and that benefits and the risk of the missed opportunities significantly outweigh the marginal direct safety and security risks.

Empirical support:

A May 2026 TELUS Digital study of 34 AI models from 10 global providers found open-source AI models are not less safe. Specifically, while open models were exploited more often on average than proprietary ones, the source of a model is not what drives risk. GLM 4.7, a large open-source model from Zhipu AI, outperformed many proprietary alternatives in safety.

Some researchers found that available evidence does not show that open models are riskier than closed models or information already available through standard research techniques such as online searching, partly because even closed models can be subverted and information for malicious actors might already be available on the internet.

2. Arguments that Open-Source is Riskier

Irreversible proliferation and guardrail removal

•The classic criticism is that open-source AI is more dangerous due to its unrestricted access, potential for malicious use, and lack of regulatory oversight.

•From an AI safety perspective, open-weight models present unique challenges. Because anyone can download open-weight models, users can often easily remove the guardrails that are meant to prevent people from using AI to hack others or build a bioweapon.

Recent research illustrates the risk:

Joint 293-day research by SentinelOne and Censys found hackers and other criminals can easily commandeer computers operating open-source large language models outside the guardrails and constraints of the major AI platforms, and direct them to carry out spam operations, phishing content creation or disinformation campaigns, evading platform security protocols.

These include hacking, hate speech and harassment, violent or gore content, personal data theft, scams or fraud, and in some cases child sexual abuse material. While some of the open-source models include guardrails, the researchers identified hundreds of instances where guardrails were explicitly removed.

In roughly a quarter of the LLMs they observed they could see system prompts, and of those 7.5% could potentially enable harmful activity.

Public perception also skews riskier: survey respondents consider open source AI tools to be riskier than proprietary AI for most types of AI risks, with the most relevant risks cited including cybersecurity (62 percent), regulatory compliance (54 percent), and intellectual property (50 percent).

3. Arguments that Closed / Proprietary is Safer

Centralized control and monitoring

•A company operating a closed model can monitor usage, patch vulnerabilities, restrict certain requests, suspend accounts, and update safety controls.

•Proprietary AI is more secure, with enterprise-grade protocols that meet compliance standards.

•Closed models often offer superior performance and are widely used in commercial applications with robust technical support and ease of implementation, though that convenience comes at cost of transparency.

This model allows centralized control over safety filters and alignment, via APIs that turn AI into a utility service.

4. Arguments that Closed is Riskier

Opacity and concentration of power

•Its closed nature limits transparency, which makes it difficult to identify and correct biases and errors.

•Proprietary, company-controlled services such as ChatGPT, Gemini, and Claude are not inherently safer: their operators can keep flaws from public scrutiny, monitor how people use them, impose restrictions, raise prices, or withdraw access.

•You might lack insight into training data, raising risks of embedded biases or unintended data retention and vendor lockin and opacity.

Secrecy also does not guarantee security. In 2023, a hacker penetrated OpenAI's internal communications and obtained information about how the company's researchers design their models, showing the threat of a breach isn't hypothetical.

5. What Regulators and Research Conclude

The most cited policy analysis is the U.S. National Telecommunications and Information Administration (NTIA) July 30, 2024 Report on Dual-Use Foundation Models with Widely Available Model Weights.

NTIA used a "marginal risk" analysis — assessing specific risks of open weights rather than general risks of AI. It concluded that the government should not restrict the wide availability of model weights for dual-use foundation models at this time and recommends continued monitoring to identify any unacceptable risks.

In recognition of these dynamics, U.S. federal policy pivoted as NTIA concluded that the marginal risks of open weights are largely speculative and recommended against immediate restrictions.

Other summaries phrase it similarly: The 2024 NTIA report framed this as a question of marginal risk and found the evidence too immature to declare open models riskier on balance, recommending monitoring rather than restriction.

The TELUS benchmark also found factors other than openness predict safety better:

•Reasoning models were significantly harder to exploit, being vulnerable to just 19.9% of attacks compared to 55.1% for models that skip the reasoning step.

•Across both open-source and proprietary models, smaller models were consistently the most vulnerable to attacks.

•Geography does not impact safety.

Bottom Line

•Open-source is potentially safer for systemic safety: bias detection, reproducibility, resilience against single-vendor failure, rapid patching by community, and preventing concentration of power.

•Open-source is potentially riskier for misuse safety: once weights are released they cannot be recalled, guardrails can be stripped, and uncensored deployments can be used for hacking, fraud, harassment and other harms at scale with no central kill-switch.

•Closed is potentially safer for containment: centralized monitoring and patching.

•Closed is potentially riskier for accountability: opacity, unverified safety claims, and dependency risks.

Most experts today, including NTIA, advocate not for banning either approach, but for monitoring marginal risks, investing in layered defenses (guardrails, prompt shielding, auditing before and after model responses), continuous adversarial testing, and shared responsibility across creators, deployers, researchers, and security teams — a model where openness enables safety research, but deployment includes controls regardless of whether the base model is open or closed.

Citations (4)

1news-articles.netLow provenance
Proprietary AI vs. Open-Source: The Battle for Control
Backs: “By keeping models closed, the industry is ignoring systemic risks that only a global community of researchers can identify and fix. From this perspective, open-source AI is actually safer because it eliminates the single point of failure and prevents a small group of engineers from having unilateral control over the…”
2akingump.comCompanies
NTIA Releases Report on Dual-Use Foundation Models with Widely Available Model Weights
Backs: “In a closed system, the "safety" of a model is determined by a private board; in an open system, the global research community can audit the model for biases, vulnerabilities, and hallucinations. Fostering innovation and reproducibility: Benefits of open model weights include fostering innovation, enhancing…”
3morningstar.comData and tools
New TELUS Digital Research Uncovers AI Safety Risks, Offers a Blueprint to Protect Enterprise AI Applications
Backs: “Specifically, while open models were exploited more often on average than proprietary ones, the source of a model is not what drives risk. GLM 4.7, a large open-source model from Zhipu AI, outperformed many proprietary alternatives in safety.”
4firm.inNot yet classified
Open-source AI models vulnerable to criminal misuse, researchers warn
Backs: “Because anyone can download open-weight models, users can often easily remove the guardrails that are meant to prevent people from using AI to hack others or build a bioweapon. Recent research illustrates the risk: Joint 293-day research by SentinelOne and Censys found hackers and other criminals can easily commandeer…”