Top 10 AI News — October 5, 2026

The weekend's biggest AI story was a security one: an open-source attack agent showed up at seven Korean financial firms, and the country's president ordered a full investigation. Elsewhere, the companies that buy AI hardest started cutting what they spend on it, while investors are still paying more for the chips underneath.

1. AI-assisted attackers breach seven Korean financial firms; President Lee orders probe

South Korean President Lee Jae Myung on Sunday ordered a thorough investigation after attackers breached seven financial institutions: Shinhan Bank, KB Kookmin, Hana, BNK Busan, Yegaram Savings Bank, Welcome Savings Bank, and Hyundai Capital. The Korea Times reported that data on more than 67,000 people was exposed, including names, contact details, resident registration numbers, annual income, and loan limits.

Investigators found the same attacker IP address at all seven firms. Local reporting named ARTEX AI, a Chinese-language open-source autonomous penetration-testing framework, alongside credential stuffing. The Financial Services Commission held an emergency meeting with bank executives the same day. FSC Chairman Lee Eog-weon warned of follow-on voice phishing and smishing, while authorities said no data that could compromise payments was leaked directly.

This is the scenario security teams have been war-gaming all year: commodity attack tooling that automates reconnaissance and exploitation, pointed at a whole sector at once. Experts quoted in Korea called for coordinated, sector-wide defenses rather than bank-by-bank security. We covered the defensive playbook in how to protect your company from agentic AI threats.

2. Meta and Microsoft scale back internal use of Claude

The Information reported that two of Anthropic's largest customers are cutting employee use of Claude. Meta's internal Claude Code user base has fallen from roughly 60,000 to about 30,000, and The Decoder, citing the report, said Meta had spent more than $105 million on Claude Code in a single 28-day period. Engineers are being moved onto Meta's in-house MetaCode and Muse Code tools.

At Microsoft, a projected Claude spend of at least $1 billion a year has been cut by more than a third. Executives Scott Guthrie and Jay Parikh have pushed staff toward GitHub Copilot and OpenAI models, and the cloud division's monthly Claude budget was reportedly reduced by more than 90%.

Cost is one driver. The bigger one is that Anthropic is now a competitor to both companies: Claude Cowork overlaps with Microsoft's office suite, and both buyers want their own coding stacks to win internally before they sell them externally. Enterprise AI spend is shifting from "use the best model" to "use our model".

3. Etched fields funding offers at $40–50 billion

TechCrunch reported Monday that inference-chip startup Etched is reviewing investment offers valuing it at $40 billion to $50 billion. The higher bids come from lesser-known investors, and talks are early. That is roughly double the $21 billion valuation Etched set when it raised $700 million in a round led by Jane Street, which came only weeks after Sequoia led a $300 million round at $10.3 billion in July.

Etched has said it holds more than $1 billion in signed customer contracts, and Jane Street is both an investor and its first customer. Etched was valued at $5 billion last December.

The bet is that inference, not training, becomes the bulk of AI compute spend, and that fixed-function silicon can beat Nvidia on cost per token there. Our analysis of the fracturing AI chip market lays out where that thesis holds and where it doesn't.

4. OpenAI begins watermarking ChatGPT text in the EU

OpenAI on Monday set out how it will meet the EU AI Act's requirement that generated text be machine-identifiable. Its textGrain system slightly alters the model's word choices to leave a detectable statistical pattern. Invisible watermarks will be added to eligible ChatGPT and Codex output for users on all plans in the EU over the coming weeks. API customers worldwide can opt in starting now, with the setting off by default.

Detector access is restricted. OpenAI opened applications on the same day, but will initially grant access only to approved researchers and expert organizations, case by case, under the EU's Code of Practice on AI-generated content transparency.

OpenAI itself described text watermarking and detection as early technologies with significant limits. The rollout matters less for what it catches than for the precedent: this is the first time a frontier lab has watermarked text by default for a whole jurisdiction, and it marks the EU AI Act's transparency rules moving from paper into products.

5. Reflection AI unveils Beam, a 501B open-weight coding model

Reflection AI introduced Beam, a sparse mixture-of-experts model with 501 billion total and 23 billion active parameters, pretrained on 23.8 trillion tokens. The company says reinforcement learning ran over four weeks on 10,500 Nvidia GB300 GPUs, with more than 100 million rollouts.

Reported scores: 77.2 on SWE-Bench Pro v2-Hard, 80.1 on Terminal-Bench v2.1, 97.8 on AIME 2026, and 90.5 on GPQA Diamond. Reflection claims reasoning scores comparable to GLM-5.2 at three to four times less inference compute. Weights, a technical report, and a model card are due later this month under Apache 2.0. Early access is through a waitlist.

Until the weights land, these are vendor numbers. If they hold up, Beam is the strongest US-built open-weight coding model, and a direct answer to the Chinese labs that have owned that tier.

6. McKinsey: AI will create more US jobs than it cuts, but 11 million workers must switch fields

A McKinsey Global Institute report, covered by Fortune on Saturday, estimates AI and automation will cut demand for about 36 million US jobs by 2035 while growth elsewhere creates about 41 million. The net is positive. The transition is the problem.

McKinsey says about 11 million workers, roughly 7% of the workforce, would need to leave their occupations entirely, with a range of 6 million to 16 million. That works out to about 770,000 people a year changing fields, around 3.6 times the historical rate. Only one in seven displaced workers has a direct path into a growing job with comparable pay and little retraining, and nearly half face paths blocked by skill or credential gaps.

Shrinking categories include office and administrative support, retail, and transportation. Growth is concentrated in healthcare, construction, and management, and McKinsey says about 76% of the new jobs cannot be done remotely. "The next decade's challenge is mobility, not scarcity," the report says.

7. Workday: demand for basic AI prompting is falling, demand for AI builders is rising

Workday's October Global Workforce Report, released Monday, found that demand for basic AI skills like prompting peaked in January 2026 and has fallen 25% since. Demand for hands-on skills, such as building AI tools, automating workflows, and AI engineering, rose 51% between September 2025 and July 2026.

Only 28% of business leaders expect AI to reduce headcount, compared with 40% who expect to get more out of the employees they already have. But internal mobility is stalling: 57% of employers saw fewer internal job moves year over year, promotions were flat, and nearly four in ten employees went through a reorganization in the past year. Most workers (79%) say they know which skills they need, but only 66% say their employer helps them build those skills.

"Employees may not be changing jobs, but their jobs are changing around them," said Workday's Phil Willburn. Prompting is now table stakes, and the pay goes to people who build with AI.

8. TikTok launches an AI shopping agent and one-click checkout

TikTok on Monday launched a conversational Shopping Assistant that answers questions about products, sizing, shipping, and availability, keeps track of a user's preferences through the conversation, and helps complete the purchase. A companion feature, Buy Direct, lets users buy from a brand with one click from the For You feed without leaving the app.

Both were built with Salesforce, Shopify, Shoplazza, and Stripe under a program TikTok calls Agentic Commerce. The strategy is defensive as much as commercial: TikTok wants to answer product questions itself before users take them to ChatGPT or another assistant.

Distribution is the advantage. TikTok doesn't need users to adopt a new shopping agent, because it can put one in front of a feed people already scroll. The payments plumbing behind moves like this is covered in the agentic commerce protocol wars.

9. Robert Half: 72% of managers will pay more for AI skills

Fortune reported Monday on Robert Half's 2027 Salary Guide data: 72% of managers plan to raise pay for recruits with AI skills, and 42% say AI expertise commands bigger salary increases than other technical skills. Separately, 57% of US managers said they are offering new hires higher salaries than they had planned.

The gap shows up in job postings, too. Average AI job postings advertise about $177,000, compared with roughly $80,000 for non-AI roles. "Employers are willing to offer higher pay for professionals who understand where AI can add value," said Robert Half's Dawn Fay.

Read together with Workday's data, the picture is consistent: employers are paying for applied AI skill, not for general familiarity with chatbots.

10. Clockwork.io raises $31M as LinkedIn and Together AI adopt GPU failover

Clockwork.io raised $31 million in a round co-led by Premji Invest, Wing Venture Capital, and Seligman Ventures, with NEA and e& Capital participating, bringing total funding to $73 million. Its LinkPass and TorchPass software keeps AI training and inference jobs running when GPUs fail, moving work to healthy hardware instead of restarting from a checkpoint.

The company says the tools are in production at LinkedIn, Together AI, and WhiteFiber. It has previously estimated that a 2,048-GPU H200 cluster could recover more than $6 million a year in compute otherwise lost to restarts.

This is a small round, but it's a real signal. As clusters grow, hardware failures become routine, and software that recovers lost GPU-hours is one of the few cost cuts available to buyers who can't get more chips.

What to watch

Whether Korea's investigation attributes the bank breaches to a single actor, and whether regulators elsewhere start treating open-source attack agents as a sector-wide threat. Reflection's Beam weights are due this month; independent benchmark runs will show whether the efficiency claims hold. And watch Anthropic's next disclosures on enterprise revenue: if Meta and Microsoft's pullback is a trend rather than two isolated cases, it will show up there first.