OpenAI has revealed that one of its forthcoming models is its most capable yet and will require further safety measures before it can be released to the public.

Dubbed Astra, the model is apparently able to discover more security vulnerabilities than any prior OpenAI model currently available, needing less computational power to accomplish this compared to other models as well.

Asta will be made available to a limited group “soon”, OpenAI said, neglecting to provide further details on which companies may be included as it touted the AI agent’s capabilities in a press conference call.

With the right tools and access, Astra can ​find previously unknown security flaws and develop ways to exploit them across many well-protected systems without ⁠a person guiding each step,” said Amelia Glaese, an OpenAI vice president supervising the safety developments.

Glaese said that the safety measures introduced may “sometimes slow, pause, or stop legitimate work” though the company is looking to minimise any disruptions to real work.

The announcement of Astra’s capabilities and the new safety measures come after industry calls to bolster security capabilities in the face of advanced AI agents and their inherent cyber risks.

Over a hundred tech firms, including those developing the most advanced AI models, called on organisations and governments to improve their security posture as they continue to develop AI models that appear capable of exploiting vulnerabilities during testing to hack into other systems.

Recommended reading

• OpenAI and Anthropic Called to White House Over AI Hacking Risks
• OpenAI Pauses AI Training Following Security Breaches
• Anthropic and OpenAI Models Implicated in New Cyber Testing Breaches
• “Sloppy” OpenAI Hack Hit More Than Hugging Face
• OpenAI Model Goes Rogue In Testing, Hacks Hugging Face

Chilling revelations from OpenAI’s Hugging Face hack showed that AI agents were capable of complex communication, coordination, and cooperation with other AI agents in order to complete tasks, and that ethical quandaries from agents did little to stop their actions.

While OpenAI revealed that the shocking hacking incident caused it to pause its AI agent development for two weeks, Astra is the first agent model that will see new safety features be introduced all together.

The firm’s safety protocol requires it to introduce more stringent guardrails on AI models if they exhibit two main characteristics: the ability to discover and exploit new cyber vulnerabilities, and plan and execute complex and new strategies for attacks. Both of these capabilities must have little to no human involvement too meet the threshold for further security measures.

Astra will now find it harder to complete potentially harmful cyber prompts, and OpenAI will continue to monitor the AI model to see if it has broken through any of these guardrails.