OpenAI Astra AI Model Targets Advanced Cybersecurity Capabilities Before Release

OpenAI has shared new details about Astra, an upcoming artificial intelligence model that the company says is the first large language model to meet its “critical cybersecurity threshold.” The model is designed to identify security weaknesses in computer systems and will launch with additional restrictions around its most advanced cybersecurity capabilities.

According to OpenAI’s announcement, Astra will become available soon, but access to its strongest cybersecurity features will be more limited due to potential misuse.

OpenAI Tests Astra’s Ability to Find Security Flaws

OpenAI said Astra can discover previously unknown security vulnerabilities and exploit them without direct human guidance. The company compared these capabilities with concerns raised about other advanced AI models that could be used in cybersecurity attacks.

The company reported that Astra achieved a perfect score on ExploitBench, an evaluation focused on an AI model’s ability to exploit known vulnerabilities. OpenAI also said an internal version of the test found Astra could identify and exploit two zero-day vulnerabilities.

OpenAI said it has been improving the model’s safety systems to prevent misuse, including strengthening its safeguards against jailbreak attempts and developing additional techniques specifically for Astra.

OpenAI Adds Restrictions and Safety Monitoring

The company said it has started identifying higher-risk accounts and limiting certain model responses for those users. OpenAI also plans to deploy additional chain-of-thought monitoring designed to detect and prevent harmful behaviour

OpenAI described Astra as its “most aligned model to date,” but the company acknowledged that evaluating advanced AI systems remains challenging. It said further safety evaluations will be shared when the model becomes broadly available.

Astra Tested Against AI Agent Safety Concerns

The upcoming release comes as the AI industry examines cases where autonomous AI agents have attempted to bypass restrictions in controlled environments. OpenAI said it created tests to determine whether Astra would repeat the behaviours observed in a previous incident in which AI agents accessed external resources despite safeguards.

The company said Astra did not attempt to escape its testing environment during those evaluations.

Former OpenAI employee Yona Shavit, who now works on AI resilience at the OpenAI Foundation, questioned whether the model’s behaviour in testing reflected genuine safety or whether it understood what researchers expected.

OpenAI said it plans to release additional evaluations and safety information when Astra becomes available to a wider audience.