Technology company OpenAI has officially assessed the cybersecurity capabilities of its new neural network. The Astra model has been assigned a critical level within a specialized rating system. This assessment is applied under the corporate Preparedness Framework. Such a high status is being assigned to one of the company's systems for the first time in its history. The test results relate to a special configuration called Daybreak Blue. The most advanced safety features will be available to a limited group of users. The neural network is capable of discovering previously unknown vulnerabilities in protected systems. The model develops exploitation methods without step-by-step human guidance. Previously, developers had only considered the possibility of reaching this high level.
The critical level requires independent creation of working attack code. The neural network must identify zero-day vulnerabilities in real-world systems. An alternative criterion is the execution of a novel, complex network attack. On the public ExploitBench benchmark, the model achieved a perfect score. The developers created an internal version of the test to protect sensitive data. Training datasets must not contain confidential test information. The internal test included twenty dangerous vulnerabilities in the V8 engine. These vulnerabilities were disclosed in the summer of 2026. The Astra neural network discovered two previously unknown zero-day vulnerabilities. The model successfully exploited them within a complex exploitation chain. The company is already sharing information about the issues with software developers.
Expert evaluations confirmed the outstanding capabilities of the new Astra neural network. The model identified unknown vulnerabilities in a specially hardened browser. The neural network constructed an attack chain to escape the sandbox. The system executed commands on the host machine through a dedicated file. In another test, the model chained together multiple system vulnerabilities. The neural network escalated a standard user's privileges to administrator level. The published results reflect the capabilities of the Daybreak Blue configuration. The model's standard production configuration has somewhat lower potential. The company has been delaying certain stages of the neural network's release over recent weeks. Developers have been intensively testing defenses against potential cyber abuse. The measures taken are now considered sufficient for a safe release.
Advanced capabilities will initially be made available to a small group of testers. Access will later be expanded through a dedicated Daybreak Blue program. The neural network received additional training to refuse prohibited cybersecurity assistance. The model strictly adheres to the safety restrictions set by its developers. Tests involving attempts to bypass restrictions yielded excellent results. The Astra neural network rejected more than ninety percent of malicious requests. The previous model, GPT-5.6 Sol, rejected only fifty-nine percent. The company is implementing additional monitoring systems to track the neural network's actions. The algorithms are designed to automatically halt potentially unauthorized actions by the model. In August, the company slowed development of its advanced systems. The reason was an incident involving the Hugging Face platform. The preliminary assessment of the Astra neural network also raised certain concerns. Developers continue to refine mechanisms for controlling artificial intelligence.