Astra Delayed Release Next Week? OpenAI Concerned About New Model's Network Attack Capability

According to Perceive Beating monitoring, OpenAI's latest test found that the upcoming Astra's programming and networking capabilities have grown so rapidly that it is now unable to rule out the possibility of autonomously attacking real critical systems. GPT-5.6 Sol was previously only in a lower tier.

By OpenAI's standards, at this tier, the model may be able to search for and exploit zero-day vulnerabilities in critical systems without the need for human supervision, and even autonomously carry out the entire process from identifying a target to designing and executing an attack.

OpenAI has suspended some of Astra's internal testing, tightened internet access, tool usage, and model weight permissions, and will next hand it over for testing to government agencies and external security organizations.

Previously leaked information indicated that Astra was set to be released next week, but at this point, that timeline seems highly uncertain. Altman also mentioned that Astra is very powerful and will eventually be open to everyone, but the risks associated with its networking capabilities will require some additional time to address.