Your question touches an active debate — I'll compare the safety trade-offs of open and closed AI.
There is no consensus that open-source is uniformly safer or riskier than closed/proprietary - evidence is mixed and experts frame it as a trade-off that depends on capability level and threat model.
Arguments that open-source is safer
•Transparency and auditability: Open models, with inspectable weights and training data, are compliant by default under regulations like the EU AI Act that require auditability, unlike black-box proprietary APIs.
•Community scrutiny: The open-source community argues "security through obscurity" is a fallacy, and open-source is actually safer because it eliminates a single point of failure and prevents a small group of engineers having unilateral control.
•Distributed red-teaming: Openness allows other people to "push and pull and prod at the models at a very at a fundamental level in order to see where might the problems be" and that community scrutiny is fundamental to spotting issues and addressing them quickly, whereas closed models rely on internal processes.
•Democratizing trustworthiness: Open-source development is argued to be preferable precisely because it disrupts the concentration of knowledge that proprietary AI creates, where few players monopolize who can investigate frontier models.
•Shared safety tooling: Open-source tools enable researchers, practitioners and policymakers to collaborate on shared risks such as algorithmic bias, adversarial attacks and lack of interpretability, and lower the barrier for under-resourced entities to adopt safety best practices.
A practical claim cited from CISA is that open source isn't inherently riskier than proprietary code — the risk comes from how you manage it.
Arguments that open-source is riskier
•Loss of control after release: Anthropic's head of international has argued open alternatives carry "more risk" of cybersecurity incidents, stating "by definition, there's more risk with them" because developers have less control over how they are used after release and that publicly available models carry greater security risks for that reason.
•Stripping of guardrails: The very transparency and modifiability that enables independent safety audits also exposes a distinct set of risks and vulnerabilities. For example, developers can locally download and modify open-source models to create versions stripped of security guardrails and safety mechanisms.
•Proliferation risk: Once weights are downloadable, they can be copied and redistributed beyond the developer's knowledge or control, and safety fine-tuning can be stripped back out. The NTIA report notes concern that open weights could make it easier for non-experts to misuse models, including toward dangerous ends.
•Misuse for high-consequence harms: Threat actors can fine-tune open models for biological weapons design, spear-phishing, non-consensual content, and other harms. The same concern extends to AIxBio: democratization of biological design tools through open-source repositories creates unprecedented dual-use risks.
•Perceived enterprise risks: Survey respondents consider open source AI tools to be riskier than proprietary AI for most types of AI risks, citing cybersecurity (62 percent), regulatory compliance (54 percent), and intellectual property (50 percent).
Closed providers counter that closed AI systems offer greater control over access and deployment, allow monitoring of usage, and help maintain security standards and prevent misuse of advanced capabilities.
What recent research finds
•Not inherently less safe: Telus Digital's second GenAI Safety Model Benchmark tested 34 AI models from 10 global providers with over 620,000 adversarial tests and found smaller models are consistently the most susceptible to attacks, regardless of whether they are open-source or proprietary. The study found open source models are not inherently less safe than closed ones, with GLM 4.7 from China's Zhipu AI outperforming many proprietary alternatives. A summary of the same study notes open source models are not inherently less safe than closed ones, with model size, reasoning capability and creator's overall safety approach as the strongest predictors of resilience. Reasoning models proved significantly harder to exploit (19.9% vulnerability vs 55.1% for standard models).
•The "marginal risk" framing: The 2024 US National Telecommunications and Information Administration (NTIA) studied models with widely available weights and framed the question as one of marginal risk — risk unique to open-weight models relative to closed models and existing technology — and found the evidence too immature to conclude that open models pose more marginal risk than benefit, recommending monitoring rather than restriction. The same finding is summarized as evidence genuinely mixed: open weights raise proliferation concerns, but openness also enables independent auditing and interpretability research that closed APIs block.
•Threshold for frontier models: For highly capable foundation models, some researchers argue that for some highly capable foundation models, extreme risks could outweigh benefits, and such models should not be open-sourced, at least not initially, while preserving open-source benefits where safe.
•Mixed incident data: Claims on safety are contested — one 2026 commentary claims 87% of privacy incidents involving LLMs came from closed models, while other analyses note closed models have also been exploited for hacking operations and that open models would likely be much easier to exploit because closed providers use automated systems to detect hacking or biological queries before they reach the model.
Conclusion: Safer in some dimensions, riskier in others
•Open-source is safer for: transparency, independent verification, avoiding vendor lock-in and single points of failure, enabling defense research ("defenders need access to models with comparable capabilities so they can detect, simulate, and respond to emerging threats"), and regulatory auditability.
•Open-source is riskier for: controlling downstream misuse, preventing guardrail removal, and preventing proliferation to non-experts for cyber or biorisk.
•Closed/proprietary is safer for: centralized control, monitoring, rapid patching and managed safety filters.
•Closed/proprietary is riskier for: opacity, concentration of power, inability of external researchers to audit, and dependence on a vendor's internal safety culture — illustrated by cases where victims had to depend on an open model to analyze a hack because closed models' security filters blocked questions about it.
Most assessments conclude openness is a spectrum — fully open source (weights + data + code), open-weight (weights only), and closed API — and policy should distinguish among them and assess marginal risk case-by-case rather than declaring one category categorically safer.