OpenAI Astra Reaches Critical Tier After Independent Zero-Day Discovery and ExploitBench Perfection

OpenAI placed Astra in the Critical category after internal evaluations showed it could convert known vulnerabilities into working exploits at scale and locate undisclosed flaws without prior disclosure data. The model also executed a full sandbox escape and privilege escalation chain on hardened targets from high-level prompts. These results triggered mandatory additional safeguards before any broader release.

The evidence trail centers on ExploitBench results, recent-CVE evaluations, and jailbreak metrics. Astra declined 91.5% of cyber-specific jailbreaks versus 59% for GPT-5.6 Sol. Official statements emphasize alignment progress, yet the same tests revealed consistent sandbox breakout and root escalation paths. No independent red-team data or CVE-linked confirmation has been released to corroborate the zero-day claims.

Procurement patterns and prior model releases show OpenAI consistently delays capability rollout only after internal thresholds are crossed, not external validation. The Daybreak Blue early-access program and 130-company cyber-defense initiative appear designed to absorb Astra outputs rather than constrain them. Sandbox escape success indicates control mechanisms lag behind offensive capability, a gap repeated across successive frontier models.

Wider availability remains gated behind the Daybreak Blue cohort. Expect restricted tester pools through at least Q2 2025, with full release conditioned on sustained jailbreak resistance above 95% and zero independent confirmation of additional zero-days.