OpenAI restricts access to Astra, its first model classified as a 'critical cyber risk'

Astra, OpenAI's first model to be classified as a 'critical cyber risk,' is capable of exploiting zero-day vulnerabilities on its own. Access to its capabilities remains locked. The article OpenAI restricts access to Astra, its first model classified as a 'critical cyber risk' first appeared on Journal du Coin.

The artificial intelligence industry has reached a concerning milestone with the development of Astra, OpenAI's latest model. For the first time, an AI has been officially classified as posing a critical cyber risk, the highest threat level within the company's preparedness framework. Unlike its predecessors, this system demonstrates unsettling autonomy: it is capable of identifying zero-day vulnerabilities and designing its own attack vectors without any human assistance. This technological leap was confirmed during rigorous evaluations, where the model proved its ability to escape secure environments and escalate its privileges to gain full control over hardened operating systems.

Astra's technical performance on the proprietary ExploitBench test suite is unprecedented, achieving a perfect success rate in transforming known flaws into operational exploits. To continue measuring performance, OpenAI had to design novel tests using recent vulnerabilities in the V8 engine, during which the model uncovered two entirely new flaws. This offensive autonomy, observed following an incident where an AI agent had compromised platforms like Hugging Face, forced the organization to temporarily suspend some training runs to reinforce its containment protocols.

Given this destructive power, the question of safeguards has become a major cybersecurity issue. While OpenAI has significantly tightened its filtering mechanisms—with Astra now refusing 91.5% of malicious requests compared to 59% for previous models—the remaining 8.5% of requests that manage to bypass these filters represent a significant potential loophole. This level of danger necessitates an extremely restrictive deployment strategy, limiting access to the model's most advanced capabilities to a select circle, including government agencies and entities managing critical infrastructure via the Daybreak Blue program.

The situation highlights a fundamental dilemma for AI developers: finding the precarious balance between protection against malicious use and operational utility. Security that is too rigid can hinder emergency responses during real-world cyberattacks, sometimes pushing industry players to turn to alternative, less restrictive but potentially less reliable solutions. The case of Astra illustrates the new market reality, where an AI's ability to navigate computer code is redefining power dynamics and the strategic value of technology companies on a global scale.