Anthropic’s GLM-5.3 Report Sparks Safety Concerns

Anthropic released a report in August claiming that the Chinese-developed GLM-5.3 model poses a major cyber-attack risk because it is freely downloadable and modifiable. The report argues that the model’s large parameter set and advanced reasoning capabilities could be weaponized by bad actors if the weights are publicly available. The claim is not a new one; it follows a long-running debate over whether frontier AI models should remain locked behind corporate APIs or be released as open-weight models.

Open vs. Closed: What the Terms Mean

A closed model, such as those powering ChatGPT or Claude, is only accessible through a web interface or API. The model’s weights and training procedure are proprietary. An open-weight model, by contrast, publishes its weights in a file that anyone can download, run locally, or fine-tune. Some open models also expose training data and hyperparameters, but the term "open-weight" specifically refers to the availability of the weight matrix.

Current Landscape of Model Performance

As of the article’s publication, the top eight models on the Artificial Analysis LLM leaderboard are all closed. Open models trail by only a few months in development, but the gap is still measurable. Closed models benefit from larger compute budgets, proprietary datasets, and continuous updates from their owners. Open models, meanwhile, offer researchers the ability to inspect internal activations, trace the chain of thought, and combine multiple models into new systems.

Why Researchers Prefer Open Models

Academic researchers need to probe a model’s internal logic to understand biases, failure modes, and potential for improvement. Closed models restrict researchers to sending prompts and observing outputs, which limits diagnostic capabilities. Open models allow researchers to load the weights into a local environment, run controlled experiments, and modify the architecture or training regimen. This flexibility is essential for reproducibility studies and for developing new techniques such as model compression or distillation.

Business Incentives for Open Models

Businesses that process sensitive data or proprietary information may prefer local deployment to avoid sending data to external APIs. Running an open model on in-house hardware can reduce operational costs and mitigate data-privacy risks. A recent Mozilla report noted that open models typically cost less than closed ones, although the report did not quantify the difference.

Democratization vs. Concentration of Power

Elizabeth Seger, senior policy advisor at the Tony Blair Institute for Global Change, argues that open models prevent mass concentration of AI power. She compares the control of frontier AI infrastructure to the control of information on the internet. If only a handful of companies can run the most advanced models, they can shape public discourse and policy. Open models could distribute that power more evenly, but they also lower the barrier for malicious actors.

The Cyber-Attack Argument

Anthropic’s claim centers on the idea that a freely available, powerful model could be repurposed for malicious code generation, phishing, or other cyber-attack vectors. The report does not provide empirical evidence of an attack but highlights the theoretical risk that the model’s capabilities could be abused. The absence of a concrete attack example means the claim remains speculative.

Trade-Offs for Policymakers

Governments face a dilemma: banning open models entirely could stifle innovation and research, while unrestricted release could increase security risks. A middle path might involve licensing, usage monitoring, or providing vetted, hardened versions of open models. The debate is currently playing out in public forums, with Nvidia CEO Jensen Huang advocating for more open models to drive hardware sales, and Anthropic’s Dario Amodei calling for caution.

What to Watch Next

The next few months will likely see more reports on the security implications of open models, as well as potential regulatory responses. Researchers will continue to benchmark open models against closed ones, and businesses will evaluate cost-benefit trade-offs for local deployment.

Sponsored

For developers looking to experiment with open models, the model hub offers a curated collection of weights and tools.

More from our coverage

The discussion of open models dovetails with recent work on Nvidia Expands Halos Safety Architecture From Autonomous Vehicles To Physical AI Hardware.

Related coverage

Explore more on this topic