OpenAI pushed agents into real payments with ChatGPT Wallet, letting "dot" agents check out on users' behalf, and launched the Daybreak security tools that use multiple agents for penetration testing and patching. LangChain raised $300 million at a $3 billion valuation, a sign investors still back a model-agnostic agent management layer. EU leaders held their first AI summit in Dublin, where Geoffrey Hinton warned of runaway AI.
Agents & tools
1. OpenAI launches ChatGPT Wallet so "dot" agents can buy things for users
On November 13, OpenAI opened ChatGPT Wallet to a limited group of US Pro subscribers: users save a Visa, Mastercard or Amex card once, and their "dot" agent picks items and pays with a token, so neither the agent nor the merchant sees the card number; each purchase is approved with Face ID or a fingerprint. It moves consumer agents into real transactions, though many retailers, including Amazon, still block AI agents; spending limits and category controls are planned.
2. OpenAI launches Daybreak Red Team and Defense Factory: agents that pen-test and write patches
On November 12, at its Cyber at the Frontier event, OpenAI introduced Daybreak Red Team (multiple agents analyze code and try to exploit weaknesses in isolated environments, constrained by a Guardian review mechanism; early access by request) and Daybreak Defense Factory (powered by GPT-6 Cyber, it turns findings into GitHub pull requests and imports results from Snyk, Endor Labs and Wiz; preview as a ChatGPT plugin). Developers must still approve patches before deployment.
3. Claude, the Claude API and Claude Code hit a widespread outage of about 2.5 hours
On November 13, from 13:09 to 15:38 UTC, claude.ai, the Claude API and Claude Code showed login and service errors; afterwards claude.ai remained degraded and scheduled tasks could run more than two hours late or be skipped. The incident is a reminder of single-provider risk for automated agent workflows. Anthropic has not published a cause.
4. LangChain raises $300 million at a $3 billion valuation
On November 13, LangChain announced a $300 million Series C led by CapitalG, with GIC, DFJ Growth, IVP and Sapphire, at a $3 billion valuation, more than double the $1.25 billion of a year earlier. The company now centers on LangSmith, a model- and cloud-agnostic platform for building and managing agents, with about 7,000 customers including Harvey, Rippling, NVIDIA, Toyota and Salesforce.
5. EU leaders hold their first AI summit as Hinton warns of "runaway AI"
On November 13 in Dublin, AI was on EU leaders' agenda for the first time, with Mistral CEO Arthur Mensch speaking on competitiveness and sovereignty and Geoffrey Hinton on societal impact. Ursula von der Leyen said Hinton outlined "enormous risks", and the talks focused on controlling frontier AI, with pharma-style pre-release testing cited as a model. The meeting was informal, so no statement or decisions were adopted.
6. Singularity-Bench: agents can now build AI training and serving systems on par with humans
An arXiv paper posted November 12 (authors including Beidi Chen, Ce Zhang, Bo Li and Ying Sheng) introduces a benchmark of six end-to-end tasks: pre-training, mid-training, post-training, inference, agent harness and cluster orchestration, compared with human-engineered baselines on the same hardware. Of 11 agents, Opus 5.5 with Claude Code leads with a Singularity-Index of 1.25, matching or beating humans on all six tasks, while the rest average 0.58. The authors also found reward-hacking holes in the Harbor framework and propose a fix called Breakwater.
7. VNPT brings AI into the administration of Hoang Hoa commune (Thanh Hoa)
According to Dan Tri (November 13), VNPT Thanh Hoa integrated AI into the management software of Hoang Hoa commune (merged from six communes, 62 officials, about 100 incoming documents a day): the AI reads and summarizes documents, splits them into tasks, assigns work, tracks progress and compiles reports for KPI reviews. Staff still check its output; it is an example of agent-style automation in commune-level administration under the two-tier government model.
8. AI Thực chiến 2026: teams build a local-heritage chatbot in 120 minutes
On November 13, Group 2 of the AI Thực chiến 2026 contest had 10 teams build a heritage notebook web app with a map, AI illustrations and a chatbot that answers in Vietnamese, cites sources and does not fabricate, on a $50 API budget (OpenAI, Google and DeepSeek models). Team AIvenger won with 953 points and advanced straight to the semifinals.
