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

This is a genuinely contested debate among AI researchers, policymakers, and industry leaders, with reasonable arguments on both sides. Let me search for the current state of this discussion to give you a well-rounded view.## There's no simple answer — it depends what kind of "safety" you mean

This is one of the most contested debates in AI policy, and reasonable experts land on very different sides. The honest answer is: it depends on which risks you're weighing, and "safety" isn't one thing.

The case that open models are riskier

Irreversibility of misuse. Once weights are released, safety controls can't be undone. As the UK AI Security Institute put it, bluntly cited by Anthropic: "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... Once open-weight models are released, these options are lost permanently: safeguards can be removed, and copies can be downloaded, redistributed, and run on private systems beyond monitoring."

Guardrails are fragile. A 2026 study testing 21 popular open-weight LLMs found they could all be tampered with despite their built-in safeguards, and the researchers concluded "the defenses available today don't yet appear strong enough to guarantee that a publicly released model will remain safe once it's in the hands of anyone who chooses to modify it."

"Abliteration" is spreading fast. A technique for stripping refusal behavior has exploded in popularity — Hugging Face currently lists over 6,000 abliterated models, compared to about 600 in 2024 — and what used to require a senior data scientist can now be done by "everybody with access to the internet and a laptop for like 400 bucks." There's documented real-world harm: one case involved an individual in a pro-ISIS chat room who claimed they used an "uncensored" AI to research explosives to destroy "Trump Tower."

Monitoring is impossible. Unlike API-based models, it's difficult to get a comprehensive picture of how people use open-weight models because they run locally and don't need the internet, so developers cannot monitor what users are asking. This matters a lot for things like bioweapons — Anthropic's discovery that Claude was misused for bio-weapons-related purposes was possible partly because it's a closed system, and that kind of oversight is harder with open-weight models.

Capability gap is shrinking. Open models used to lag well behind, but UK AISI found the cyber-capability gap has narrowed from 6 to 10 months in 2025 to just 4 to 7 months — meaning dangerous capabilities reach unrestricted open models faster than before.

The case that open models are safer (or at least not uniquely dangerous)

Transparency enables scrutiny. Open models let people audit code, inspect data sources, and test for biases — though that same openness exposes vulnerabilities in plain sight, inviting security experts to find and fix flaws but also giving attackers the same blueprints. Closed models simply require trusting the vendor.

Industry pushback. Over 70 companies, including major cloud/chip providers, argued against banning open weights: more than 70 companies including Google, Microsoft and NVIDIA said open-weight models make advanced AI more accessible, and while acknowledging "real and distinct risks," argued open models "broaden defensive capability, increase transparency, and allow vulnerabilities to be discovered and remediated," adding that defenders need comparable capabilities to detect and respond to threats used by attackers.

"Vaccine" argument. Some researchers argue forced exposure to misuse strengthens collective defenses over time — open-source AI can be seen as a "vaccine" for human society, since its potential for malicious use forces society to strengthen defenses, whereas if only closed, aligned AI existed, people might grow complacent, and failure or theft of that closed AI could be catastrophic.

Concentration-of-power risk. Many don't see this as a safety question at all but a power question: "The clearest, biggest argument [for open models] is around preventing mass concentration of power—it is about making sure that this foundational technology is a public good," says Elizabeth Seger of the Tony Blair Institute. Anthropic's own CEO even noted the most dangerous model might actually be one trained in secret and handed only to a military or security apparatus for drones and surveillance — i.e., closed-source concentrated in the wrong hands can be scarier than open diffusion.

The emerging middle ground

Even leading safety-focused labs reject the binary framing. Anthropic explicitly states: "A blanket ban on open-weights models is neither the correct remedy... All sufficiently capable models, open and closed, should go through mandatory safety testing." Researchers increasingly argue capability level matters more than open/closed status — a SaferAI analysis noted "the frontier of capability is not the frontier of risk," so mitigations must be assessed alongside raw capability.

Bottom line

So the most defensible answer: open-source AI is riskier on misuse-containment dimensions, but potentially safer on power-concentration and transparency dimensions — and the two sides are often answering different questions when they claim one is "safer" than the other.