OpenAI's GPT-6.2 Astra jumps 22 points on ExploitBench, ships tighter cyber safeguards
JiaweiLiu_ · x · 2026-11-18
OpenAI officially announced GPT-6.2 Astra as its most capable cyber model ever, jumping 22 points over GPT-6 Astra on ExploitBench Internal Port (success rates reaching 53%). Given the increased capabilities, OpenAI is launching Astra and Sol with tighter safeguards for high-risk cybersecurity assistance and bringing them to Daybreak so approved defenders get reduced refusals.
More from Models
- Microsoft ships MAI Code 2 and Code 2 Flash coding models into GitHub Copilot — marlene_zw · 2026-11-18
- Matt Shumer pans a just-released model: "no idea why they released this" — daniel_mac8 · 2026-11-18
- Matt Shumer: this model release needed 'a lot more iteration' — mattshumer_ · 2026-11-18
- GPT-6.2 Astra early review: best-in-class 3D, but frontend styling needs babysitting — flavioAd · 2026-11-18
- GPT-6.2 Astra scores first-ever 100% on ARC-AGI-3 via provider adapter — mark_k · 2026-11-18
- OpenAI launches GPT-6.2 Astra and Sol; early user recreates cross-section book visuals — craigsdennis · 2026-11-18