Why it matters
The difficulty was not that commercial models lacked capability. Hugging Face said their safeguards could not distinguish an incident responder submitting authentic attack commands from an attacker seeking the same technical help. Running GLM 5.2 locally allowed the team to analyze large volumes of sensitive evidence without sending credentials or attack artifacts to an outside API.
That does not make every open-weight model safer, nor does it settle the policy debate over access. It shows that cyber governance must account for both sides of the contest: attackers are not bound by provider policies, while defenders need audited escalation channels, locally controlled tools, and plans established before a crisis.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
CNBC — How a Chinese AI model stopped an OpenAI cyberattack OpenAI — Security incident during model evaluation Associated Press — What the OpenAI and Hugging Face incident revealed
