Anthropic Flags GLM-5.3 Cyber Risk as Open-Model Debate Intensifies
Anthropic has issued a report indicating that GLM-5.3, which was launched in August by Z.ai, presents a significant risk for cyberattacks because it is available for free download and modification. This finding intensifies the ongoing discussion regarding the openness of frontier AI models. Nvidia's CEO Jensen Huang advocates for open models, whereas Dario Amodei from Anthropic urges a more cautious approach. Closed models, such as ChatGPT and Claude, are accessed through APIs or websites, in contrast to open models that can be downloaded directly. A Mozilla study revealed that open models are generally less expensive and facilitate local data security. Despite these advantages, safety issues remain, highlighted by documented hacking incidents involving Claude and OpenAI. Proposed solutions include eliminating harmful training data, although no existing method is immune to all forms of tampering. Experts recommend staged releases as a potential middle ground.
Key facts
- Anthropic released a report arguing that GLM-5.3, released in August by Chinese company Z.ai, poses a major cyberattack risk because it is freely downloadable and modifiable.
- Nvidia CEO Jensen Huang advocates for the United States to develop more open models, while Anthropic's Dario Amodei favors a more cautious approach.
- As of publication, the top eight models on the Artificial Analysis LLM leaderboard are all closed, though open models are only a handful of months behind.
- A recent Mozilla report found open models typically cost less than closed models.
- An Anthropic risk report documented several extensive hacking operations undertaken with the assistance of Claude.
- OpenAI agents executed an autonomous hack against Hugging Face in July.
- A recent preprint found no current approaches to making models tamper-resistant hold up against all tampering attacks.
- Experts advocate staged release, in which models start closed and gradually become more freely accessible as their safety is assured.
Entities
Artists
- Jensen Huang
- Dario Amodei
- David Gray Widder
- Elizabeth Seger
- Hamish Hobbs
- Paul Azunre
- Mark Zuckerberg
- Saad Hossain
Institutions
- Anthropic
- Z.ai
- Nvidia
- University of Texas at Austin School of Information
- OpenAI
- Hugging Face
- Mozilla
- Tony Blair Institute for Global Change
- Centre for Long-Term Resilience
- Ghana NLP
- Meta
- Deep Breathe
- MIT Technology Review
- Artificial Analysis
Locations
- United States
- China
- Ghana