Daily AINo. 16第 16 期
Agents are now an API
Safety reports and agent plumbing moved faster than the weekend news cycle.
Anthropic’s threat report: AI is already running parts of the kill chain
Anthropic’s September 2026 threat intel covers Dec 2025–Aug 2026 misuse across cyber ops, influence campaigns, surveillance, fraud, bio, weapons, and distillation. The through-line: actors used Haiku/Sonnet/Opus as orchestrators (recon, phishing, malware rebuild loops), not just chat helpers — and Anthropic says it disrupted the campaigns and shared intel with partners.
SourcesAnthropic · Sept 2026
OpenAI ships Agents API (public beta)
OpenAI put the Codex-style harness behind a public Agents API on Sept 10 — managed sessions, tool use, subagents, plus hosted or bring-your-own sandboxes. No separate Agents API fee; you pay for tokens, tools, and (if you use them) hosted sandbox containers.
SourcesOpenAI · Sept 10, 2026
ChatGPT for Financial Services goes live
OpenAI launched a finance-flavored ChatGPT for banks and investment firms, built with Morgan Stanley and Evercore and wired to licensed data from S&P Capital IQ, LSEG, PitchBook, Moody’s, and more. TNW flags the missing EU data-residency / DORA detail that European buyers will still have to negotiate.
SourcesThe Next Web · Sept 11, 2026
Washington pressure rises; OpenAI flips toward mandatory rules
After Anthropic researcher Jacob Coxon resigned warning labs are racing irresponsibly — with Evan Hubinger backing him — more US lawmakers are pushing new AI rules. Same window: OpenAI’s Chris Lehane urged Congress to pass compulsory frontier safety laws before December, a clear shift off “voluntary commitments only.”
SourcesThe Japan Times · Sept 11, 2026Mathrubhumi · Sept 11, 2026
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
OpenAI把Codex套件开放了，安全监管也在升温
周末最值得盯的，不是又一个模型名字，而是：智能体开始「可产品化交付」，安全与监管也同时升温。
Anthropic 威胁情报：AI 已在帮黑客跑流程
Anthropic 发布 2026 年 9 月威胁情报报告，覆盖 2025 年 12 月到 2026 年 8 月、七类滥用场景（网络攻击、舆论操控、监控、诈骗、生物、武器、蒸馏等）。重点不是「AI 会写恶意代码」，而是威胁方把 Claude 当成编排器：侦察、钓鱼、改马、并行代理一起上；Anthropic 称已打断并与合作方共享情报。
OpenAI Agents API 公测：Codex 那套 harness 对外开放
9 月 10 日，OpenAI 把驱动 Codex 的 harness 做成 Agents API 公测：管会话、工具调用、子代理，还能选官方托管沙箱或自带环境。没有单独的 Agents API 附加费，按 token、工具，以及（若用）托管沙箱计费。
ChatGPT 金融版上线：模型 + 行情数据一起卖
OpenAI 推出面向银行与投行的 ChatGPT for Financial Services，和摩根士丹利、Evercore 一起打磨，并接入 S&P Capital IQ、LSEG、PitchBook、Moody's 等授权数据。欧洲买家还会追问数据驻留和 DORA——TNW 指出官宣里几乎没写清楚。
美国监管升温，OpenAI 转向「要强制规则」
Anthropic 研究员 Jacob Coxon 辞职并警告实验室在不负责任地竞速，Evan Hubinger 公开附和后，更多国会议员呼吁新规。同一时段，OpenAI 全球事务负责人 Chris Lehane 呼吁国会在 12 月休会前通过强制性前沿安全立法——等于从「自愿承诺够了」翻盘。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作