OpenAI: Astra is on its way – a new model ‘capable of rejecting 91.5 per cent of unlawful requests’
The new OpenAI model has also been introduced in the wake of the incident last July, when over a thousand AI agents went out of control during an internal security test and breached the Hugging Face platform’s systems
1' min read
1' min read
“Astra has been fully trained for some time now and represents a significant step forward,” announced Sam Altman, president of OpenAI, the parent company of ChatGPT, in a post on X .
Astra will introduce ‘stricter security measures’, aimed at ‘more reliably blocking malicious IT requests and complying with security requirements, providing additional safeguards against misuse, and implementing monitoring systems capable of blocking potentially unauthorised activity’.
According to the parent company of ChatGPT, 91.5 per cent of unlawful requests are rejected.
The new OpenAI model has also been introduced following an incident last July, when over a thousand artificial intelligence agents escaped control during an internal security test and breached the systems of the Hugging Face platform.
This incident is not an isolated case: models from the company AI Anthropic had also gained unauthorised access to three organisations – the names of which have not been disclosed – during tests that were supposed to keep them isolated.
“Astra represents a significant advance in cybersecurity capabilities,” explains the blog of OpenAI, which is why it is the first programme to be classified as ‘critical’ under the Preparedness Framework. For the time being, however, access to the most advanced features will be restricted to a group of testers.
