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

Open vs. closed AI safety — I'll lay out the risks and safeguards on both sides.

It is not categorically safer or riskier - open-source and closed/proprietary AI have different risk profiles, and which is "safer" depends on what kind of safety you mean.

Arguments that open-source AI is SAFER

1. Transparency and auditability:
Open-source principles are argued to foster accessibility, collaboration, responsibility and interoperability. Compared to proprietary software, open-source benefits include being free to inspect and modify, developed by a diverse community, and errors being detected and corrected much faster than in closed-source software.

Specifically for AI:

A Linux Foundation GenAI survey found proprietary solutions are not perceived as more secure than open source alternatives, challenging the traditional notion that closed systems inherently offer superior security. Similarly a ZDNet analysis described it as Too close to call, noting open-source is audited by many eyes which helps find holes.

2. Secrecy does not equal security:
A 2024 review in the journal Science including Princeton researchers concluded that limiting public access to AI models could stifle innovation by restricting engineers' ability to improve models, increase secrecy, and concentrate power in a few corporations.

They found available evidence does not show that open models are riskier than closed models or information already available through search engines. This is partly because even closed models can be subverted and partly because information for malicious actors might already be available on the internet. Researchers argue regulators should assess marginal risk: To what extent do open foundation models increase risk relative to closed models or to preexisting technologies such as search engines.

3. Avoids concentration of power and black-box risks:
Exclusive proprietary AI increases concentration of power, restricting access and hampering innovation, fair competition and scientific progress. Proprietary systems are difficult for the public to appraise, and thus to identify and fix errors. Without open-source models, the public will know less and large tech companies will have more influence over design and execution.

Nature Computational Science argues that nevertheless the benefits are likely to outweigh the downsides of open-source AI, especially if a responsible open-source approach with clear barriers against misuse is pursued.

Arguments that open-source AI is RISKIER

1. Guardrails can be easily and irreversibly removed:
Unlike proprietary API-based models such as Claude or ChatGPT where safety protections operate server-side and cannot be modified by end users, open-weight models can be modified once downloaded.

A joint investigation by the Financial Times and AI safety group Alice, published May 2026, demonstrated that a free tool called Heretic hosted on GitHub can strip all safety protections from open-weight models including those from Meta, Google and OpenAI in under ten minutes using only a standard laptop. Tests found guardrails on models from Meta and Google could be removed in under 10 minutes without specialist hardware, after which modified systems responded to prompts the originals refused, including requests linked to malware and chemical hazards.

The scale is growing: Hugging Face, which hosts open-source models, currently lists over 6,000 abliterated models compared to about 600 in 2024. Once released, open models cannot be recalled or centrally patched.

2. Supply-chain and unpredictable risks:
The growing openness fosters transparency and faster iteration, but AI models are still software bundled with extensive codebases, dependencies and data pipelines and can harbour vulnerabilities, outdated components or even hidden backdoors that scale with adoption.

AI models are black boxes built from massive, opaque datasets and hard-to-trace training processes. Even when datasets are available they are often too large to audit. Malicious behaviours can be trained in and the non-deterministic nature makes exhaustive testing impossible. There is no definitive way to prove that an AI model is safe, reliable, or free from unintended consequences.

Other cited societal risks include use for development of weapons and AI-generated propaganda campaigns.

3. Current perception of closed as safer:
For now, closed models lead in capability and are perceived to be safer. From a cybersecurity perspective, closed-source AI presents a paradox: while its restricted access reduces surface-level risks such as tampering and unauthorized use, it can also create blind spots.

Even closed models are not immune: a Cisco study of 15 closed/proprietary flagship models found single-turn attack success is not a reliable proxy for attacker adaptation across turns, and benchmarks may give an insufficient basis for safety decisions.

Consensus view

Most researchers and policy analyses now reject a simple binary:

For that reason proposals for responsible open-source AI call for layered access controls similar to sensitive open datasets like MIMIC-III which is open only after ethics training, novel licenses such as RAIL licenses which prevent irresponsible uses by granting permission only for certain use cases, plus governance, visibility into AI usage, supply-chain scanning and continuous monitoring.

In short: open-source does not automatically democratize or secure AI, and closed-source does not automatically make it safe. Risk depends on governance, not just openness.