Amir Golestan

Capability 8: He showed him the cybersecurity classification that made OpenAI gate its own model. He told him GPT-6 Astra is the first OpenAI model to receive a "Critical" cybersecurity classification under the company's Preparedness Framework. "Critical" means the model's cybersecurity capabilities are significant enough to require access controls beyond the standard release. Specifically, Astra scored 100% on ExploitBench a benchmark testing whether the model can turn known software vulnerabilities into working exploits. GPT-5.6 Sol scored 78.5%. The jump from 78.5% to 100% isn't a performance improvement. It's saturation the model exploits every known vulnerability in the benchmark without a single failure. He told him OpenAI's response was to gate the advanced cybersecurity capabilities. The public Astra model the one available to ChatGPT Plus, Pro, Business, and Enterprise subscribers refuses to generate proof-of-concept exploits, write attack code, or assist with offensive security operations. The ungated model available only to vetted organizations through OpenAI's Daybreak trusted access program operates with reduced safeguards for legitimate defensive security research, penetration testing, and vulnerability assessment. He told him the cybersecurity gating is a preview of how frontier AI models will be released going forward. The model that can defend networks can also attack them. The model that can identify vulnerabilities can also exploit them. The capability is dual-use by nature. OpenAI's response restricting the offensive capabilities while releasing the defensive capabilities widely is the balance every AI company will navigate as models become more capable. He told him for the average ChatGPT user, the cybersecurity classification is invisible the public model simply refuses certain requests. But the classification signals something important about the model's overall capability: GPT-6 Astra is powerful enough that OpenAI delayed its release by 4 weeks, gated its most advanced features, and built new safety evaluations specifically to ensure the model doesn't exceed its intended scope. The model is released, but it's released on a leash. The previous models didn't need one.