OpenAI 与 Anthropic 一周动态：GPT-6 Cyber、Images 2.6 与 Claude 个人理财功能
一周两家大动作：GPT-6 Cyber 复现上千个内核漏洞，Claude 也能管你的银行账户了，还有开源的零知识证明工具。
OpenAI 发布面向网络安全的 GPT-6 Cyber，自主流水线用它复现了 1724 个 Linux 内核漏洞，并为 948 个提权漏洞写出可用 exploit。ChatGPT 周活跃用户两个月内从 10 亿增长到 12 亿，Images 2.6 新增 GIF 制作与参考图保存功能。Anthropic 在美国以测试版推出 Claude 个人理财功能，通过 Plaid 以只读方式接入银行、卡与投资账户。Anthropic 还开源了主要由 Claude 构建的零知识证明系统 FlashProve，用于在不接触权重的情况下验证响应来自特定承诺模型。METR 独立评估认为 Anthropic 能力提升速度约为此前两倍。
OpenAI and Anthropic this week: GPT-6 Cyber, Images 2.6, Pierre, FlashProve, Personal Finance, METR (Week 45, 2026)
OpenAI introduced GPT-6 Cyber, a cyber-specialized variant of GPT-6 Astra for select Daybreak Red customers, rated Critical in cybersecurity under its Preparedness Framework
An autonomous OpenAI pipeline running GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and wrote working exploits for 948 privilege escalations
ChatGPT Images 2.6 adds GIF creation and saved reference images, and the new Flare and Sunburst API models generate images 50 percent faster
OpenAI is acquiring Pierre Computer Company to build Git workflows for coding agents into Codex
OpenAI told Big Technology that ChatGPT weekly active users grew from 1 to 1.2 billion in two months
A Family Basic lineup sits in ChatGPT's web checkout pricing config, hinting at an unannounced basic family tier
Anthropic introduced Personal Finance in Claude as a US beta, linking bank, card, loan and investment accounts through Plaid with read-only access and no training on those chats
Anthropic's Messages API now has message threads in public beta, so conversations live server-side and each request sends only new messages
Anthropic open-sourced FlashProve, a zero-knowledge proof system built largely by Claude to verify a response came from a committed model without seeing its weights
Anthropic's Frontier Red Team had Claude play EVE Online, where newer models traded profitably, and EVE Frontier, where a secretly sabotaging agent was almost never caught
METR's independent assessment finds Anthropic likely improving capabilities on well-specified tasks at about twice its earlier pace, with much of the most informative evidence redacted, and more
