• Home
• Technology

Sam Altman, CEO of OpenAI, revealed that the artificial intelligence model "GPT-6 Astra" was not the prospective project the company recently signaled it had halted, clarifying that the training of "Astra" had already concluded some time ago.

Speaking in an interview with Bloomberg, Altman stated that during testing, "GPT-6 Astra" reached what he described as a "critical" threshold in cybersecurity capabilities. This performance milestone prompted OpenAI to significantly reinforce safety measures and protective guardrails before making the model accessible to users.

Altman explained that the model's advanced proficiency in cybersecurity compelled the company to manage its deployment under newly calibrated safety protocols. Internal evaluations demonstrated that "Astra" had become capable of executing complex offensive and defensive cyber operations at a level demanding additional layers of oversight, alignment, and operational constraints.

Altman's statements reflect growing concern among frontier AI labs regarding advanced models attaining autonomous proficiencies in identifying vulnerabilities, reverse-engineering systems, and executing sophisticated cyber tasks. While these tools offer transformative value to enterprise defenders and security researchers, they introduce acute dual-use risks if exploited maliciously.

The disclosure surrounding Astra’s capabilities comes as leading AI developers compete to deploy more autonomous, complex problem-solving architectures, alongside an escalating focus on red-teaming and pre-deployment safety evaluations. Altman emphasized that the new safeguards are a direct response to the model's observed thresholds, underscoring that frontier AI development is no longer defined solely by compute scaling and raw benchmarks, but by the continuous reassessment of systemic risks emerging with each generational leap.