GPT-6 Astra is the new security model from OpenAI, presented within the Daybreak program. The company claims it has reached a critical level of cybersecurity.
Additionally, OpenAI asserts that GPT-6 Astra allows for discussions about artificial general intelligence (AGI) due to its capabilities for learning, reasoning, and performing cognitive tasks.
OpenAI claims that GPT-6 Astra enables discussion of artificial general intelligence (AGI)
What GPT-6 Astra Can Do
According to OpenAI, the model can detect unknown security vulnerabilities and develop new ways to exploit them.
The model can also perform these tasks on protected systems without human supervision, according to the company's statement.
OpenAI indicated that Astra is more robust than GPT-5.6 Sol. In particular, it better withstands attempts to remove its restrictions, known as jailbreak.
What can GPT-6 Astra do
It also offers greater resistance to message injections or prompt injection and exhibits behavior more aligned with expected instructions.
The New Security Measures of GPT-6 Astra
The company strengthened security measures during the training and implementation of GPT-6 Astra.
The new security measures of GPT-6 Astra
To develop these protections, OpenAI incorporated lessons from the hacking of Hugging Face, executed by autonomous AI agents powered by the company's models.
The model also incorporates new protective measures for minors under 18. If it detects interaction with a minor, it applies additional limits on various types of content and activities.
Sexual content.
Eating disorders and risks related to body image.
Age-restricted products.
Dangerous activities.
Emotional dependency.
Self-harm and graphic violence.
Furthermore, GPT-6 Astra cannot engage in romantic role-playing games or promote dependency relationships that could replace real-world connections.
OpenAI incorporated the lessons from the Hugging Face hack
GPT-6 Astra Improves in Mental Health and Hallucinations
OpenAI also highlighted improvements over GPT-5.6 Sol in areas related to mental health, emotional dependency, and self-harm.
The model also shows fewer errors related to hallucinations and a lower likelihood of reproducing hallucinated information, according to the company.
GPT-6 Astra improves mental health and hallucinations
Who Can Use GPT-6 Astra
For now, GPT-6 Astra is available to business partners participating in the Daybreak program.
OpenAI plans to later launch a commercial version for paid users of ChatGPT, with additional security measures.
OpenAI Claims GPT-6 Astra Achieves AGI
The president of OpenAI, Greg Brockman, stated that GPT-6 Astra qualifies as artificial general intelligence, according to Bloomberg and The Washington Post.
OpenAI claims that GPT-6 Astra achieves AGI
The concept of AGI refers to an artificial intelligence capable of learning, reasoning, and performing intellectual or cognitive tasks at the level of a person or even surpassing it.
Brockman maintained that GPT-6 Astra represents “a real shift” in the type of tasks that people can delegate to AI and in how it can enhance their capabilities.
