OpenAI Unveils Astra, First AI Model Rated ‘Critical’ for Autonomous Zero‑Day Exploits

OpenAI announced that its newest model, Astra, can independently discover previously unknown software vulnerabilities—so‑called zero‑day flaws—and generate functional exploits without human input, marking the first time the company has classified an AI system at the highest, “Critical,” cyber‑risk tier.

The company said Astra’s ability to locate and weaponize vulnerabilities is fully autonomous, meaning the model can scan codebases, identify security gaps and produce working exploit code without external guidance. By reaching the “Critical” designation, Astra becomes OpenAI’s most hazardous cybersecurity model to date, prompting the firm to place it under strict internal controls and limited external access.

OpenAI’s disclosure follows a statement made in August, in which the organization acknowledged it could not rule out the possibility that an upcoming model might possess such capabilities. The latest reveal confirms those concerns, showing that the company’s own risk assessments have moved from speculative to concrete.

Industry experts note that the emergence of an AI that can autonomously generate zero‑day exploits raises the stakes for both defenders and attackers. Traditional vulnerability research relies on skilled analysts manually probing software, a process that can take weeks or months. An AI that can accelerate this timeline dramatically could flood the market with new exploits, potentially overwhelming existing patch‑management practices.

Security researchers have long warned that generative AI could be weaponized, but Astra represents a tangible milestone. The model’s capabilities could be misused by criminal groups or state actors seeking to compromise critical infrastructure, financial systems, or proprietary technology. At the same time, defenders may eventually harness similar tools for defensive purposes, such as automated vulnerability discovery to accelerate patch cycles.

OpenAI says Astra will remain under “tightest possible restrictions,” with access limited to a small internal team and subject to rigorous monitoring. The company also indicated that it is working with external partners and policymakers to develop safeguards, though specific mitigation strategies were not disclosed.

The announcement adds to a growing debate over how AI developers should be regulated when their products cross the line into high‑risk cybersecurity domains. Legislators in the United States and Europe have recently introduced proposals aimed at imposing safety reviews and licensing requirements for AI systems capable of generating malicious code. Astra’s debut may accelerate those discussions, as regulators grapple with the challenge of balancing innovation against national security concerns.

For now, the cybersecurity community is watching closely to see how OpenAI manages Astra’s deployment and what steps are taken to prevent its misuse. The model’s existence underscores a broader shift: AI is no longer just an aid for defensive analysis but is becoming a potential autonomous actor in the cyber‑threat landscape.