South Korea Probes Signs That AI Was Used to Hack Its Banks

South Korea's president, Lee Jae Myung, said Tuesday that "signs have emerged" that AI models were used in recent cyberattacks on the country's banks, "causing considerable public concern and anxiety," Laura Chung and Jin Yu Young report in The New York Times. "It's now become possible to use A.I. to hack with ease even without specialized skills," he told a livestreamed cabinet meeting, ordering officials to establish what happened "swiftly and clearly." It is the first suspected major banking hack conducted with AI.

Seven financial institutions have reported breaches since Sept. 30: Shinhan, KB Kookmin, Hana and BNK Busan banks, Yegaram and Welcome savings banks, and Hyundai Capital. The Record and Quartz put the total exposed at about 68,000 people. Shinhan said the personal credit information of about 25,000 people was leaked, including incomes and loan limits; Yegaram Savings Bank estimates about 40,000 customers. No money is known to have been taken from customers' accounts. The worry is what comes next: the stolen details are exactly what a phone scammer needs to pose convincingly as a victim's bank. The government has issued a consumer alert, and the banks have promised to cover losses that stem directly from the leaks, without yet saying how.

Investigators found traces of Artex AI, a Chinese-language, open-source tool on GitHub that uses AI to find weaknesses in computer systems; its creator built it to help organizations test their own networks, Quartz reports. Officials stress that this does not identify the attackers, who routed their traffic through some 20 internet addresses in more than 10 countries, including the United States, Japan and Germany. "It is impossible to identify an attacker based on an IP address alone," said Park Sang-won, head of the Financial Security Institute. No one has been named.

The breaches have also stalled a policy shift. Korea requires banks to keep internal systems walled off from the internet, and regulators had begun granting exemptions so banks could run AI tools to defend themselves. On Tuesday the Financial Services Commission postponed the next round, after breaches hit even banks it had judged to have strong security, including Shinhan and Hana, which were in the first round.

Sources: The New York Times — Laura Chung and Jin Yu Young · The Record · Quartz · The Korea Herald · AFP via Malay Mail · Asia Business Daily — regulation postponed · The Korea Times — consumer alert

Why it matters: D.A.D. recently covered a modeling study warning that banks' shared dependence on a handful of AI vendors could turn a single hack into a system-wide event (D.A.D., September 10). This is the other half of the picture: the attackers. If freely available AI tools let people without specialist skills break into banks, every institution holding customer data faces more attackers than it planned for. And Korea's answer, letting banks use AI to defend against AI, has just been put on hold.

Source: nytimes.com

Claude Now Works Inside Google Docs, Sheets and Slides: How to Use It

Anthropic has put Claude inside Google Docs, Sheets and Slides. A new add-on, in beta on all paid Claude plans, opens Claude in a sidebar next to the file you're working on. It reads what you have open, including any text, cells or slides you've selected, and edits the file directly.

- Docs: fixes a sentence or restyles a heading without disturbing the formatting around it. For bigger rewrites, it proposes changes as cards you can apply or dismiss. - Sheets: writes formulas, builds pivot tables and charts, and adds tabs. For heavier data cleaning, it runs the numbers in Python and writes the results back into the sheet. - Slides: builds new slides that match the deck's existing layouts and theme, then flags text that overlaps, runs off the slide or is hard to read.

By default Claude asks before every edit; an "Accept all edits" mode lets it work through a task without stopping. The sidebar carries over your existing Claude connectors and skills, so it can, for instance, pull a client's account history from Salesforce into a deck.

How to start: install "Claude" from the Google Workspace Marketplace, open a file, then go to Extensions > Claude > Open Claude. Separately, new connectors let you create and edit Google files from inside Claude itself; its access matches your existing Google sharing permissions. On Team and Enterprise plans, an administrator has to turn the connectors on first.

Sources: Anthropic · Claude Help Center — setup · Google Workspace Marketplace

Why it matters: For the many offices that run on Google, this ends the routine of copying text into a chat window and pasting the answer back. Anthropic is not first here: Google's own Gemini is built into Workspace, OpenAI put ChatGPT into Sheets in April and let it edit Google files from the chat in August, and Claude has been generally available in Microsoft Word, Excel and PowerPoint since May. It also comes three weeks after Anthropic launched its own Docs and Slides to compete with Google's. Now it is meeting Google's users where they already work.

Source: claude.com

OpenAI Posts AI-Generated Math Proofs Publicly, Skipping Peer Review

OpenAI published a batch of new mathematical results generated by an internal frontier model, posting them to GitHub with formal proof verifications (Lean) rather than through peer-reviewed journals. The company says it consulted outside mathematicians at the Institute for Advanced Study on how to responsibly release AI-derived proofs, and disclosed compute costs—on average, the equivalent of three hours of ChatGPT Pro usage per result. Some mathematicians online welcomed the open release as more transparent than paywalled journals; others said OpenAI acted only after public criticism and that real transparency questions remain unresolved.

Sources: OpenAI · Discuss on Hacker News

Why it matters: AI systems are starting to produce original mathematical work fast enough that the bottleneck shifts from discovery to verification and trust—forcing the math community to build new norms for checking claims nobody can yet peer-review the old way.

Source: openai.com

Mistral Previews a Trillion-Parameter Open Model, Holding Back the Weights for Cyber Testing

French lab Mistral released a preview of Mistral Large 4, a 1.05-trillion-parameter model it says it will release as open weights, free to download and run, by the end of October. For now it is available only through Mistral's paid service. Like Reflection's Beam this week, it is pitched as an alternative both to closed American models and to Chinese open ones, and it was trained in Mistral's own European data centres. It is unusually strong at cybersecurity, solving 93% of the challenges on Cybench, a set of 40 security exercises. Before the weights go out, Mistral is testing it with "cybersecurity leaders, vetted partners, and state authorities," who get a version with reduced moderation and expanded cyber capabilities.

Sources: Mistral · TechCrunch · Discuss on Hacker News

Why it matters: Western open models are arriving in a rush, and this one gives Europe its own. But skill at security work cuts both ways, and once an open model's weights are released, anyone who downloads them can strip out its safeguards. The Korean bank attacks, which police are tracing to an open-source AI hacking tool, show why that matters.

Source: mistral.ai

Jira and Confluence Get OpenAI-Powered Assistants to Plan Your Work

Atlassian is deepening its OpenAI partnership, bringing OpenAI's frontier models into Jira, Confluence, and its Rovo AI assistant to power agents that plan and execute work. The companies say the models will tap Atlassian's "Teamwork Graph" — the web of project data, tickets, and docs companies already have — so agents can act with context on who's doing what. More than 3,000 Atlassian developers already use OpenAI's Codex internally, and OpenAI itself reportedly runs core operations through Jira.

Sources: OpenAI

Why it matters: If you manage projects in Jira or Confluence, expect AI agents that can draft tickets, flag blockers, and summarize work automatically rather than just answer questions about it.

Source: openai.com

OpenAI Says GPT-6 Astra Can Run Days-Long Research and Handle Contract Work

OpenAI published two customer write-ups this week for GPT-6 Astra, its current flagship model. At quant trading firm Jump Trading, Lucas Baker, who heads LLM research and development, says Astra lets agents take on loosely defined work, from building entire services to analyses that run for days, that used to need frequent human guidance. Separately, OpenAI said Astra is its first frontier model trained on tasks from Ironclad, a contract-management software company. On 11 tasks built with Ironclad to mirror what contracting teams do, such as configuring agreements, managing approvals and handling renewals, Astra averaged 55%, against 41.6% for its predecessor, GPT-5.6 Sol, while cutting average time per attempt from 37 minutes to about 19. An internal model used during Astra's development scored 63.7%. The scores measure how well agents met task requirements, not the share of jobs completed.

Sources: OpenAI — Jump Trading · OpenAI — Ironclad

Why it matters: Both are company-published case studies, not independent tests. But they show where OpenAI is aiming: long stretches of unsupervised work in finance and law. That is the same territory where, late last month, it pulled GPT-6.1 Astra for being too willing to act without permission.

Source: openai.com

AI Tool Helps NYC Students Find High Schools Without Overcrowding Top Picks

Researchers built an AI recommendation tool to help NYC high schoolers find nearby, high-performing programs they're likely to get into, then tested it in a randomized trial during this year's admissions cycle. A naive version of such tools can backfire: if too many applicants get steered toward the same popular schools, acceptance odds actually drop, hurting students with the fewest nearby options most. The team's "congestion-aware" system avoided that trap—applicants who got recommendations were 57% more likely to rank a suggested program and 71% more likely to match into one, with no one rejected from a recommended school.

Sources: arXiv

Why it matters: It's a concrete example of how recommendation algorithms—the same mechanics behind Netflix or Amazon suggestions—can worsen inequality in high-stakes public systems like school admissions unless designed to account for the fact that everyone's getting the same advice.

Source: arxiv.org

Where Companies Sit in AI Hiring Networks Boosts Their Value

A new study mapped how AI talent moves between companies, drawing on roughly 535 million employment records across 58 countries from 2010 to 2022. The finding: it's not just how much AI talent a firm hoards that boosts its value, but its position in the broader talent network—who it hires from and loses people to. AI talent clusters around leading firms more intensely than general hiring does, and companies that gain central network positions see enterprise value rise afterward, even accounting for size and assets.

Sources: arXiv

Why it matters: For executives building AI teams, this suggests that where your hires come from—and who's poaching your people—may be a more telling signal of competitive strength than headcount alone.

Source: arxiv.org

Study Finds Drivers Quietly Smooth Over Ride-App Algorithm Errors

A qualitative study of an on-demand ride-pooling service finds that when algorithmic routing or pricing decisions clash with what passengers expect, it's drivers who absorb the fallout—smoothing over confusion, apologizing for the app's choices, and managing frustration in real time. Researchers call this "Frontstage Mediation Work": labor that keeps automated systems looking seamless but goes unrecorded in performance metrics, logs, or job descriptions. The study identifies four recurring practices drivers use to paper over these gaps, though it offers no quantitative data on how often this occurs.

Sources: arXiv

Why it matters: As companies push more decisions onto algorithms—routing, pricing, scheduling—the human workers interfacing with customers quietly inherit the job of managing the system's mistakes, a cost that rarely shows up on any balance sheet.

Source: arxiv.org

Study Finds Most Proposed AI Safeguards for Kids Remain Untested

A review of 100 studies on children and teens interacting with AI—in schools, homes, and public services—found a wide gap between identifying risks and actually fixing them. Researchers say most proposed safeguards exist only as ideas or prototypes, never implemented or tested in real settings. Even when countermeasures are evaluated, studies tend to measure whether the technology works as designed rather than whether it actually protects kids from harm.

Sources: arXiv

Why it matters: As schools and parents rapidly adopt AI tools for children, this suggests the safety claims behind them are running well ahead of any real evidence they work.

Source: arxiv.org