Friday, 9 October 2026
Seven Banks, and a President Saying It Takes No Skill Anymore
South Korea Thinks AI Agents Broke Into Seven Banks — and the Part That Should Worry You Is the President’s Phrasing
At least seven South Korean financial institutions were breached, among them Hana Bank, KB Kookmin Bank and Shinhan Bank, and on Tuesday President Lee Jae-myung said officials believe AI agents were behind at least some of it: “Signs have emerged” they were deployed. The data of at least 68,000 people is out — borrowing history, incomes, phone numbers, names — with Shinhan alone putting roughly 25,000 customers in that number. The incidents came to light on September 30. The Financial Supervisory Service has traced 33 IP addresses across at least twelve countries, Japan, the US, Thailand, Vietnam and Hong Kong among them; the National Office of Investigation has put 28 investigators on it; officials suspect a Chinese tool called Artex AI without having confirmed it. All of that is the ordinary shape of a bank breach story. This is the line that isn’t: “It’s now become possible to use AI to hack with ease even without specialized skills.” A head of state saying the barrier to entry is gone is a different kind of disclosure than a bank saying it lost 25,000 records. And the pattern has a history now — Australia said last month that OpenAI agents got into its Medicare database, and OpenAI disclosed in July that its own agents were behind a hack of Hugging Face.
Source: The Record, October 6, 2026
Anthropic Named Its Defence Programme and Then Said Some Fixes Might Take Decades
Two days after widening who may use its strongest models for cyber defence, Anthropic gave the whole effort a name on Thursday: the Anthropic Cyber Mission. New since Tuesday’s access tiers are two concrete things. The Critical Infrastructure Defence Programme pairs frontier Claude models, on-site engineers and Anthropic’s threat research with eleven founding partners — Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation — aimed at the operational technology behind power grids, water systems and transport. And OSS Scanner is a free, opt-in scanner for enrolled open-source projects, modelled on Google’s OSS-Fuzz, which sends each report with a proof of concept and a suggested fix. Anthropic expects a true-positive rate above 90% and says plainly that “the reports are model-generated and sent without human review.” The candour is the striking part. Of its earlier Project Glasswing pilot, now folded into the verification programme, the company writes that partners found plenty of vulnerabilities but “we haven’t yet achieved a sufficient reduction in cyber risk”; months routinely passed between finding a hole and closing it; and because industrial machinery cannot be taken offline, a fix “may have to wait until it can be applied safely to running machinery — in some rare cases, this might take decades.” Finding the bug was never the hard part.
Source: Anthropic, October 8, 2026
Eleven Companies Pledged $2.4 Billion to US Science, and Almost None of It Is Money
The White House announced on Thursday that eleven industry partners have committed $2.4 billion to the Genesis Mission Consortium, the federal AI-for-science effort now spanning more than fifteen agencies. NVIDIA leads at $1 billion, then AMD at $500 million, OpenAI at $200 million, Anthropic and Google at $150 million each, AMP and Emerald AI at $100 million each, and AWS, Armada, Crusoe and Micron at $50 million each. The fact sheet’s own wording is the thing to read closely: “$2.4B in SI for science tools and compute credits.” Not cash — tools and credits, which is to say each company’s own product, valued at its own list price, given to researchers who will use it on that company’s hardware. That is a real contribution and it is also a customer-acquisition programme, and both of those can be true at once. Anthropic’s $150 million, which it describes as three years of Claude, Claude Code and API credits across several hundred projects at NASA, NIH, NSF and others, is the one commitment with a published breakdown. The round sits inside a larger set of announcements — over $6 billion all told, including a $100 million fellowship for four-year PhDs and a $215 million quantum competition — and follows a $5 billion federal commitment in July.
Source: The White House, October 8, 2026
OpenAI Will Sell You Speed for Six Times the Price, Which Makes the Cheap Model Dearer Than the Flagship
Ultrafast began rolling out on Thursday for GPT-6.1 Sol, in the API, in Codex and in ChatGPT Work — “near-Astra intelligence at up to 8x faster speeds,” by OpenAI’s own description, available in all supported regions with US and EU data residency, and on the Pro 500, eligible usage-based Enterprise and credit-based Edu plans, where an administrator has to switch it on. The pricing is where it gets interesting. Ultrafast costs $12 per million input tokens and $60 per million output, against the $2 and $10 Sol has carried since it launched on September 30. That is six times the standard price of the model whose entire pitch was costing a fifth of flagship GPT-6 Astra — and by OpenAI’s own note, about 1.2 times what Astra itself costs. So the cheap model, run fast, is more expensive than the expensive model run normally. Nothing about the model changed; latency is now a separately priced product, chosen per request rather than per account. Which tells you what the bottleneck has quietly become for anyone running agents at volume: not whether the thing can think, but how long you are willing to sit there while it does.