OpenAI ships GPT-6.2 Astra: 22-point ExploitBench jump, tighter cyber safeguards
charliermarsh · x · 2026-11-18
OpenAI officially released GPT-6.2 Astra and Sol, its most capable cyber models yet, alongside tighter safeguards for higher-risk cybersecurity assistance.
- GPT-6.2 Astra jumps 22 points over GPT-6 Astra on ExploitBench Internal Port; researcher team itszn13 reports a 22% increase on unseen v8 bugs, another leap akin to 5.6→6
- Charts show ExploitBench success rates exceeding 50% at high token budgets; on DeepBinary (Windows binary vuln finding) GPT-6.2 Astra racks up >$1M estimated bounty at $4k cost, while GPT-6.2 Sol is most cost-efficient at low budgets; SecureSWE reaches 65%
- Large improvements also in secure code generation and complete, secure patch generation
- Models are coming to Daybreak so approved defenders get reduced refusals
Related event: GPT-6.2 Astra Boosts Exploit Skills, $12 Per Vulnerability(2 posts)→
More from Models
- OpenAI ships 6.2 Astra and Sol, immediately panned as its worst model release — Hesamation · 2026-11-18
- GPT-6.2 draws backlash on X, but early hands-on feels faster even at compaction — CtrlAltDwayne · 2026-11-18
- ChatGPT 6.2 models panned: "vibes are bad off the charts" — minchoi · 2026-11-18
- Australian couple built a 220B sovereign AI model on under $10M, ranked 43rd worldwide — aakashgupta · 2026-11-18
- Users say Opus 5.5 refuses near-impossible tasks while rivals at least try — imjustnewatai · 2026-11-18
- User slams Grok for lacking automatic troll-muting on X — robleclerc · 2026-11-18