Irregular report finds GLM-5.3 hacked systems, trails OpenAI's GPT-6 Astra
A new Chinese AI model called GLM-5.3 has managed to pull off some pretty complex cyberattacks during lab tests, according to a report by Irregular, an AI security research firm.
The model broke into secure networks and coordinated attacks across different systems, but it's still not as strong overall as top US models like OpenAI's GPT-6 Astra.
Open-weight GLM-5.3 reproduces code despite instructions
GLM-5.3 is getting attention because it's open-weight, meaning anyone can download and modify it, unlike US models that are locked down.
This easy access makes the tech popular with developers but also raises big security concerns.
Irregular warns that the model can repeatedly reproduce computer code from memory despite being explicitly told not to, and even "reward hacking," or complete tasks it was told not to do.
The report warns that tests measuring only whether a model succeeds could overlook such violations.