GPT-6.2 Astra/Sol Spotted in Google Cloud Console; Cyber Model Nailed 948 Kernel Exploits

新智元 · wechat · 2026-11-06

The Information reports OpenAI may ship GPT-6.2 Astra, Sol and Cyber as soon as this Friday; the models have already appeared in a Google Cloud console quota page. Leaks suggest Astra targets cheap, high-frequency use with five tiers ($0.44–$11.05 per task, 14%–59% scores); Sol uses the same 'game training' and high-intensity environments as the strongest internal model rather than being a mere distillation; the internal Cyber model is extremely expensive and strong—its Low tier ($11/task, 76%) reportedly beats Astra's Max tier.

Alongside, OpenAI published a report on a fully autonomous vulnerability-research pipeline powered by the Cyber model targeting the Linux kernel: a five-stage agent pipeline (scout, analyst, triage, exploit dev, patching) consumed $1.213M in API credits in a single run, reproducing 1,724 vulnerabilities and producing working local privilege-escalation exploits for 948 of them ($1,279 each, versus $20K–$70K typical bounties), with patches disclosed to maintainers per community guidelines. Sampling found the AI invented at least 10 novel exploitation techniques unseen by senior experts, and even found bugs in the QEMU test VM without access to its source code. Notably, a stronger model given no hints outperformed one hand-fed 103 human techniques.

The report warns that within a few years frontier AI will inevitably find any vulnerability that exists, making high-assurance core software critical in 'a world without 0-days.' (Release details are leaks, unconfirmed.)

Original post →

GPT-6.2 Astra/Sol Spotted in Google Cloud Console; Cyber Model Nailed 948 Kernel Exploits

More from Models