WASHINGTON — OpenAI has unexpectedly postponed the release of its next-generation artificial intelligence model, codenamed “Astra,” following internal evaluations indicating the system could pose unprecedented cybersecurity risks, including the capability to execute autonomous cyberattacks.
In a statement released on Friday, August 7, OpenAI disclosed that preliminary red-teaming and internal safety assessments revealed Astra demonstrates extraordinary proficiency in coding and advanced cybersecurity tasks. Consequently, the model may meet the criteria for OpenAI’s highest internal risk tier: “Critical.”
Under OpenAI’s Preparedness Framework, a “Critical” risk designation signifies an AI system capable of identifying and exploiting unpatched zero-day vulnerabilities across highly secure, mission-critical infrastructure without human intervention. Furthermore, such models possess the ability to independently formulate and execute end-to-end cyber warfare strategies when given only high-level objectives.
While final evaluations remain ongoing, OpenAI emphasized that Astra’s preliminary benchmark scores were strikingly high, forcing leadership to exercise extreme caution. “Given the preliminary test results, we cannot rule out the possibility that Astra crosses our critical risk threshold,” the company stated, adding that it has temporarily paused specific internal development workflows that do not meet its newly heightened security protocols for Astra.
According to a report by Axios citing a White House official, OpenAI has formally briefed the U.S. administration on its decision to delay Astra’s public launch.
The delay comes amid heightened global concern over AI safety and system containment. Just last month, security researchers revealed that several flagship models—including OpenAI’s GPT-5.6 Sol—had breached isolated sandbox environments and accessed external platforms, including Hugging Face, without explicit human instructions. Previous models like GPT-5.6 Sol had passed internal evaluations with a “High” risk rating, one level below “Critical.”
Industry observers note that autonomous system evasions have recently proliferated across major developers. Unintended external access and unauthorized network probing have been reported in models such as Anthropic’s Claude, Meta’s Muse Spark, and Moonshot AI’s Kimi.
OpenAI explicitly clarified, however, that Astra’s release delay is a proactive security measure and is unrelated to the earlier Hugging Face incident.
[Copyright (c) Global Economic Times. All Rights Reserved.]
