OpenAI Says ‘Astra’ Can Independently Construct Sophisticated Attacks

OpenAI says its forthcoming Astra model has reached the company’s highest cybersecurity capability category, becoming its first AI system to receive a “Critical” designation.

According to OpenAI, Astra can detect previously unknown software vulnerabilities and build working exploits from them without a human directing each individual step. The capability represents a significant shift toward more autonomous cyber operations.

The company introduced the classification in a Tuesday post, saying Astra is the first model to meet the “Critical” threshold outlined in its Preparedness Framework.

To reach that level, a model must be able to independently locate zero-day vulnerabilities and create exploits capable of working against hardened real-world environments. The criteria also cover systems that can turn a broad attack objective into a functioning cyberattack with little or no human assistance.

Astra produced strong results during OpenAI’s evaluations. It achieved a perfect score on a benchmark testing whether it could develop exploits based on known vulnerabilities. During another internal test, it discovered two previously unknown flaws while assembling an exploit chain.

OpenAI said the model also managed to escape a hardened browser sandbox and execute commands on the host computer. In a separate experiment, Astra identified several operating-system vulnerabilities and combined them to obtain root access.

The company has slowed elements of Astra’s development while implementing additional safeguards. Its most advanced cyber capabilities are expected to remain restricted to a select group of testers initially.

The development is particularly relevant to cryptocurrency security, where a successful exploit can rapidly turn a software vulnerability into a financial loss. AI systems capable of automatically analyzing code, spotting misconfigurations and linking vulnerabilities could dramatically shorten the time required to construct an attack.

Security researchers have emphasized that the more immediate concern is not necessarily the creation of entirely novel attack methods. Instead, AI could make existing vulnerabilities much easier to discover and exploit at unprecedented speed.

Astra’s progress also fits into a wider trend among frontier AI systems. Models are increasingly demonstrating the ability to perform complex technical and reasoning tasks with less direct human involvement, moving beyond traditional chatbot interactions and basic code-generation capabilities.