AI Hacks Korean Banks, Haiku Gets Cheap, Windows Goes Agentic

AI hacking tool tied to South Korean bank breaches

CrowdStrike says a likely single, Chinese-speaking attacker breached multiple South Korean financial institutions between late September and early October using ARTEX, an open-source AI penetration-testing tool posted on GitHub in July that automatically hunts for security flaws. The tool ran on models including DeepSeek v4.1-flash, GLM-5.3, and Grok 4.6, and researchers found Claude Code session logs on the attacker's open directories showing searches for Telegram groups to sell stolen data. Korean newspaper Khan reports more than 25,000 customer records, including names, contact details, income, and credit limits, were taken from Shinhan Bank alone. Source: The Decoder

South Korea's financial regulator held an emergency meeting, and President Lee Jae Myung called for a thorough investigation. The New York Times notes CrowdStrike has not named an individual or group behind the attack. Anthropic separately found that GLM-5.3 can write exploits nearly on par with its own frontier model, Mythos Preview, the one that touched off the AI-capability debate in late March 2026. Sources: The Decoder, New York Times - AI

The breach, the stolen records, and the tool itself are all documented, making this a real incident rather than hype. What is unproven is attribution, and whether the attacker needed any skill beyond knowing how to run a GitHub repository. Sources: The Decoder, New York Times - AI

Anthropic slashes Haiku pricing, closing gap with OpenAI

Anthropic has released Claude Haiku 5.5, cutting list prices for prompts under 100,000 tokens by about 90 percent to 10 cents per million input tokens and 50 cents per million output, down from $1 and $5 for Haiku 4.5. Factoring in a new, slightly hungrier tokenizer, Anthropic's own average savings estimate comes to about 75 percent, and the new pricing matches OpenAI's GPT-6 Luna exactly at that size. Sources: SiliconANGLE - AI, MarkTechPost, Simon Willison

On Anthropic's own benchmarks, Haiku 5.5 beats Luna on every test where both have a score, including 72.4 percent against 48.9 percent on OSWorld 2.1, a test of agents operating a real computer through long tasks. Anthropic also halved cache-read prices on Sonnet 5.5, and Simon Willison reports that the company added monthly API credits for Max and Team subscribers, up to $500 for Team accounts. Sources: SiliconANGLE - AI, MarkTechPost

The release is real but mostly a price story rather than a capability leap. Anthropic itself still points customers to Sonnet or Opus for serious agentic coding work. Source: SiliconANGLE - AI

Microsoft lets Copilot reach into your files and act on its own

At its first live event in two years, Microsoft unveiled Hybrid Intelligence, a mix of local and cloud models that lets Copilot reach into a user's files and take actions across Windows rather than just answer questions. In an onstage demo, an executive asked the Autopilot agent to gather tax documents for an accountant, and it searched folders, renamed files, zipped them, and drafted the email itself. Source: The Verge - AI

Those file-access features are set to roll out over the next couple of months, while a related Windows search update that lets users trigger quick actions, like toggling dark mode from the search bar, arrives this fall. Source: The Verge - AI

Nous Research hits $1.5 billion valuation on agent revenue

Nous Research, maker of the open-source Hermes agent, has raised a $90 million Series B led by Robot Ventures, with Nvidia and Samsung among the participants, valuing the company at $1.5 billion. Hermes has been downloaded or cloned more than 24 million times and, by Nous's own estimate, drives about 2.5 percent of global AI token usage. Sources: SiliconANGLE - AI, TechCrunch - AI

Hermes is a desktop agent that can browse the web, run code in a container, and organize files, built on models like Claude plus Nous's own Hermes 4 series. Nous was at roughly $36 million in annualized revenue by mid-September 2026 and expects to pass $100 million before the end of 2026, and the new funding will support its push into the enterprise sector with Hermes for Businesses, which lets companies deploy customized AI agents for multi-step workflows while keeping their data private and secure. Sources: SiliconANGLE - AI, TechCrunch - AI

Revenue, a lead investor, and a named valuation all check out, making this a real raise rather than a paper round. Sources: SiliconANGLE - AI, TechCrunch - AI

Google opens its AI watermark detector, with limits

Google has opened SynthID Detector to everyone on a dedicated website that spots the invisible SynthID watermark, and Google says Gemini alone has integrated SynthID with over 180 billion images and videos plus 240,000 years of audio. Users log in with a Google, OpenAI or Apple account and get roughly ten checks a day, a limit Google engineers say is meant to stop people reverse-engineering a way around the watermark. Sources: SiliconANGLE - AI, The Decoder, Ars Technica - AI

The tool has clear gaps. Grok does not use SynthID, nor do many open-weight Chinese models, so the detector has nothing to find in that content, and it cannot say which part of an image was AI-edited versus real, only whether a watermark exists at all. Google says on the site it is still, in its words, continuously working on extending its partnerships. Sources: SiliconANGLE - AI, The Decoder, Ars Technica - AI

It is a genuine, useful tool for a real slice of AI content, but much AI-generated media that lacks a SynthID watermark will not be flagged at all. Sources: SiliconANGLE - AI, Ars Technica - AI

Meta's Muse agent arrives on iPad a month after launch

Meta's Muse agent has landed on the iPad a month after its phone debut. Sensor Tower estimates Muse has topped 6.6 million installs since its early-September launch. Sources: The Verge - AI, TechCrunch - AI

With the iPad release, Meta added new connectors including Canva, Dropbox, Figma, GitHub, Notion and Granola, along with an update that lets small businesses give Muse a business goal the agent then uses to plan marketing and client pitches. Sources: The Verge - AI, TechCrunch - AI

OpenAI's shopping agent stumbles on a simple couch purchase

A Wired reporter let OpenAI's always-on agent, Dots, help buy a couch, and the attempt did not go smoothly. The agent, available through ChatGPT for $100 a month, got the reporter's name wrong immediately, tried to solve a captcha it could not handle, and at one point blurted out "oh I love you too, Reece" after apparently mishearing a phone call. Source: Wired - AI

In the author’s account, the agent still did useful work, searching couch options against real dimensions and budget and sending back product links. The story shows this agent can handle some real browsing tasks but also makes enough mistakes that users need to supervise it closely. Source: Wired - AI

New motion dataset tests humanoid robot, but results are unclear

A new benchmark called HiPHI offers a 617.5-hour whole-body human motion dataset captured with sub-millimeter accuracy, including 245.7 hours of people handling objects, with policies trained on it tested on a physical Unitree G1 humanoid robot. Source: IEEE Spectrum - AI

The evidence confirms the dataset's scope and that trained policies were deployed on the real robot, but it does not say how the robot actually performed, leaving the practical payoff an open question. Source: IEEE Spectrum - AI