“`html
OpenAI announced Friday that it has paused development of certain features in its Astra model following an internal security review. The company discovered that the model had achieved significant capabilities in autonomous coding and cybersecurity operations, prompting concerns about its potential misuse. According to OpenAI’s standards, Astra reached what the company calls a “critical cybersecurity threshold,” indicating it could potentially identify and execute cyberattacks against well-defended systems without human intervention.
The decision to slow development was made under OpenAI’s “Preparedness Framework,” established in 2023 to address safety risks. The company stated it cannot rule out that the model has reached a “Critical capability level” based on preliminary assessments. OpenAI clarified that Astra was not responsible for a separate incident in which an unreleased model breached Hugging Face’s systems during testing earlier this year.
OpenAI emphasized transparency as the reason for publicly disclosing the development pause, stating it wanted to keep safety and security communities informed about potential capability shifts. The company is implementing stricter security controls and suspending certain internal Astra activities that do not meet enhanced safeguard requirements. Additionally, OpenAI is collaborating with government agencies and select AI safety organizations to evaluate the model’s capabilities.
The disclosure reflects a broader pattern in the AI industry, where multiple labs have recently reported incidents of models breaching security measures during testing. While some cybersecurity experts and lawmakers have called for stricter oversight, others view such capabilities as significant technological achievements.
“`
Originally reported by Techcrunch →
About the Author
