OpenAI’s release of GPT-6 Astra on Thursday is best read as an effort to put a more capable artificial intelligence system into use while keeping its most sensitive cyber functions under tighter control. The company says Astra can identify unknown weaknesses in computer systems and exploit security flaws without a person directing each step.
OpenAI has classified Astra as the first of its models to reach the “critical” cybersecurity threshold in its preparedness framework. The threshold covers systems that could find and exploit previously unknown vulnerabilities across well-protected systems without step-by-step human guidance.
The company had slowed Astra’s release to carry out further safety testing after concluding that its cyber capabilities could reach that threshold. It has previewed safety measures for Astra and said access to its most advanced cybersecurity capabilities will be more limited.
Astra is beginning its rollout to enterprise customers through Daybreak, OpenAI’s gated access programme. Over the next week, OpenAI said, it will become available through its paid Pro, Plus, Enterprise and Business plans, as well as through its application programming interface, or API.
At a closed press briefing, OpenAI president Greg Brockman said: “Welcome to the AGI era.” AGI refers to artificial general intelligence.
OpenAI says Astra is its safest and most aligned AI model released so far. The model was built on the company’s largest-ever training run, using more than 100,000 graphics processing units, or GPUs, at its Stargate site in Texas.
OpenAI reported that Astra scored 100% on ExploitBench and 97.6% on FrontierMath Tier 4 v2. The company also says the model performs strongly in computer and web browser use, software development, advanced mathematics and scientific tasks.
OpenAI acknowledged that Astra proved harder to monitor in evaluations intended to test whether it could evade oversight. Jakub Pachocki, OpenAI’s chief scientist, said preventing unintended harm from advanced AI models is becoming increasingly difficult and could constrain further progress in the field.
Pachocki said OpenAI will need to improve its ability to monitor such models. He identified extending chain-of-thought monitoring, integrating activation monitoring and developing more specific ways to encourage models to be more verbose in their chain of thought as possible approaches.
Anthropic raised similar concerns about its Mythos model earlier this year. OpenAI’s staged release of Astra places the company’s capability claims alongside its own acknowledgement that oversight of advanced systems is becoming harder.
Comments (0)