South Korea Plans to Inspect Every Glasses-Wearing Student at Its College Exam

The Korea Times reports that South Korea’s Ministry of Education plans to require exam supervisors to inspect every test taker wearing glasses before each session of this year’s College Scholastic Ability Test, known as the CSAT or Suneung, which is about a month away. Possessing AI smart glasses, even without using them, will count as cheating. Violators will have their results invalidated and be barred from taking the exam the following year.

The rule follows recent incidents at universities. On October 10 a student at Yonsei University was caught wearing smart glasses after reportedly switching from ordinary glasses once the test had begun. On October 3 the University of Seoul removed a student wearing smart glasses during an essay exam and barred the applicant for three years.

Detection is the hard part. Radio-frequency detectors are unlikely to be in place for this year’s exam, Education Minister Choi Kyo-jin said at a parliamentary audit, because differences between testing sites could cause detection errors. The ministry plans training videos to help supervisors tell AI glasses from ordinary eyewear. Teachers’ groups say that isn’t enough and have petitioned for extra inspection staff and clear procedures for disputes.

CoreWeave’s Debt Fears Spread Through AI Stocks, and Huang Says Nvidia Won’t Backstop It

CoreWeave shares fell about 15% on Wednesday, according to market reports, and slid another 3.3% on Thursday to around $63.59 as investors worried about debt-funded AI buildouts. Its bonds were hit as well, and rival neocloud providers including Nebius, Applied Digital and Iren also declined. CoreWeave is widely treated as a stand-in for how risky the whole AI trade has become.

Dow Jones, citing Bloomberg, reports that Nvidia CEO Jensen Huang said his company doesn’t need to backstop CoreWeave because its finances are improving and its cost of capital is decreasing. Nvidia is CoreWeave’s supplier, customer and investor, and Reuters reports it has agreed to buy unsold CoreWeave capacity through April 2032. Propping CoreWeave up further could also feed accusations of circular financing, where Nvidia’s balance sheet keeps its own customers spending.

AI Weekly, citing Bloomberg, adds that Huang told an event in Berlin on October 16 that the selloff will pass and that the AI buildout is still at the very beginning, as chip stocks fell for an eighth straight session, the longest streak since 2008. I couldn’t independently confirm the Berlin remarks or the streak length. For more on how AI infrastructure is being financed, see our October 14 AI news roundup.

AI-Assisted Hack Leaks About 100,000 Records From 16 Korean Financial Institutions

Security Magazine reports that an AI-assisted attack on the financial sector has leaked roughly 100,000 personal records across 16 known institutions in South Korea. They include four banks (Shinhan, KB Kookmin, BNK Busan and Hana), four savings banks, and credit finance firms, online investment companies, electronic financial service providers and cryptocurrency businesses, among them Hyundai Capital and KDB Capital. Five of the institutions were recently identified as exposure victims.

The leaked data includes names, email addresses, contact information and resident registration numbers. Security executives quoted in the report say AI lets an attacker work through an entire sector at a speed that used to be reserved for a single target, and that defenses built on recognizing known-bad behavior will struggle against attackers who change methods with each attempt. An earlier Security Magazine report links the campaign to an AI-powered penetration-testing tool.

Reflection AI Releases Beam, a 501B Open-Weight Model Built to Use Fewer Tokens

Reflection AI released Beam on October 16 under the Apache 2.0 license, according to VentureBeat as summarized by AI Weekly. It’s a mixture-of-experts model with 501 billion total and 23 billion active parameters, aimed at coding, agents, science and math. The company says it scores about 80% on Terminal-Bench 2.1 while using roughly 25,000 tokens per attempt, compared with about 80,000 for GLM-5.2 at a similar score. Those are Reflection’s own figures.

Reflection says it trained Beam for four weeks on 10,500 Nvidia GB300 GPUs with more than 100 million reinforcement-learning rollouts that penalized unnecessary tokens, and that the weights fit on an eight-GPU server in 8-bit or 4-bit precision. If the token efficiency holds up, it targets agent costs rather than raw price per token. I haven’t verified the benchmarks. For a look at an agent-first coding tool, see our Cursor AI review.

ChatGPT’s Chrome Extension Can Now Drive Your Open Tabs

OpenAI’s ChatGPT extension for Chrome now has two modes, according to 9to5Google as summarized by AI Weekly. In Chat mode it can read across your open tabs to gather information. In Work mode it can actively control them: send messages, fill in forms, download files and group tabs. The extension reportedly has about 7 million users.

The update moves OpenAI’s computer-use features into Chrome itself, following the discontinuation of its Atlas browser. It’s another sign that always-on, action-taking agents are becoming standard, a theme we explored in our Dots by OpenAI review. As with any agent that can click and type for you, review what it does before approving anything that sends, buys or deletes.