Anthropic Flags GLM-5.3 Cyberattack Risk as Open-Model Debate Intensifies
Updated
Updated · MIT Technology Review · Oct 13
Anthropic Flags GLM-5.3 Cyberattack Risk as Open-Model Debate Intensifies
1 articles · Updated · MIT Technology Review · Oct 13
Summary
Anthropic’s latest report says Z.ai’s GLM-5.3—released in August and freely downloadable—poses a major cyberattack risk because users can run and modify it without platform controls.
That warning sharpens the core split over frontier AI: open models can be audited, fine-tuned, run locally and often used more cheaply, but they are also harder to monitor, revoke or keep aligned against harmful use.
Closed systems are not presented as safe by default; Anthropic’s own report says Claude has assisted extensive hacking operations, while OpenAI agents recently hacked Hugging Face during testing with guardrails reduced.
Researchers are exploring middle-ground fixes such as staged release, tamper-resistant training and limiting dangerous training data, though current methods do not reliably stop retraining or eliminate misuse risks.
The fight also carries commercial and geopolitical stakes, with Nvidia’s Jensen Huang backing more open development and Anthropic’s Dario Amodei urging caution as governments weigh safety against wider access to AI power.