Source : BUSINESS NEWS
EARLIER this month, the US government restricted non-American citizens from using Anthropic’s Fable and Mythos artificial intelligence models.
US-based Anthropic announced the Mythos model in 2026 but was concerned it was too powerful to let anyone use it. The company feared it could find and exploit cybersecurity holes in all major software systems.
So they initiated ‘project Glasswing’ in April as a way to help get big, trusted companies’ software cyber-ready before public release.
Project Glasswing was initially open to other large American companies, and eventually other trusted government allies, with the Australian government among these.
There was a lot of hype around this model but reports indicated it was very capable. Some were sceptical, as this was not the first time an AI company had announced concern about the dangers of releasing their model to the world.
In 2019, for example, Open-AI withheld its GPT-2 model because it was too dangerous. Yet here we are, seven years later, with GPT-5.5 in open access and the world hasn’t collapsed in an insurmountable pile of AI slop (at least not yet).
As pressure mounted on Anthropic to release its Mythos model, it announced a restricted version named Fable. Controversially, Anthropic’s safeguards analysed the user’s input and silently switched the model behind the scenes to its less-capable Opus 4.8 model.
Shortly after launch, however, users complained that the safety filters were too eager, and people with legitimate requests felt as though they had been silently switched. Anthropic later apologised for this approach but then the US government export restrictions came into effect and locked the model down.
This had the effect of removing access to non-US citizens in the earlier project Glasswing initiative.
Anthropic’s hype was too effective in scaring the US government.
On the same day the US government locked down Fable (June 13), Chinese AI lab Z.ai released the best open-weight AI model to date, GLM 5.2.
Early benchmarks show this model is almost as good as Opus 4.8, which means about as good as the dumber model Fable would reroute to when receiving a suspect request.
Unlike Opus 4.8, the GLM model is entirely open source, with the structure and the weight available for anyone to use. Anyone with access to a super computer or a data centre can run it now. And you can run it and be confident it won’t route your request to an even dumber model.
That’s not to say Chinese models don’t have their own quirks. Many of these models will put forward the perspectives of the Chinese government when asked certain topics, but since they are open, developers can adjust their internals to make them more willing to do anything the user asks.
This means they can also be coerced to help with cybersecurity exploits or other bad things. And this is a real risk.
As these models get better and better they can be used to help create even better models, known as a ‘fast takeoff’ artificial general intelligence (AGI) scenario.
However, I think that, in a fast takeoff scenario, it’s better for the AGI to be shared and open than closed and locked down.
If you don’t care about any AI research stuff and just want to know if the Chinese models are safe to use, the answer is yes. And if you are really paranoid, then the safest way is to host the model yourself.
If you are working with sensitive data or software, setting it up so it runs on your own servers and with restricted or no internet access will give you near cutting-edge AI at a fraction of the cost and with 100 per cent privacy.
• John Vial has a PhD in robotics and has spent the past several years leading teams in major Perth businesses focused on AI and robotics
