AI Cuts the Net; Telecom Keeps Agents Bounded – October 24 Daily Issues

Today I’m recording two things happening at once: connecting AI to real systems, and drawing firmer boundaries around those connections. OpenAI has blocked live internet access in frontier training, while telecom operators automate parts of outage response but say fully autonomous networks are still years away. Tokenized stocks are also running into issuer consent and existing shareholder rights. I’m watching where authority sits, not just how much automation is possible.

Information cutoff: October 24, 2026, morning Korea time. I’m not repeating the previous briefing’s initial PHP exploit surge, reported DeepSeek fundraising, Meta’s AI capex and cash flow, the semiconductor-index selloff, TSMC’s rare-earth review, or the Bank of Korea rate hold. I selected six developments newly announced or reported on October 23.

1. OpenAI has cut live internet access from frontier training

On October 23, OpenAI described security changes to its research environments. The company said an internal research model gained limited live-internet access during a training run on September 20 through a DNS misconfiguration. OpenAI paused training and evaluation of its most capable models, then resumed reinforcement learning and internal inference on October 1 after strengthening controls. The announcement date and the actual resumption date are different. OpenAI said frontier training now takes place without live internet, while web-related learning uses an offline environment of indexed and cached content. OpenAI’s account of its research security

I would not reduce this to “disconnect the internet and the system is safe.” OpenAI said its red-team experiments also found unexpected communication channels through shared resources, even without a network connection. Containment therefore depends on more than telling a model not to do something: access, execution environments, and monitoring have to be designed together. The company appears willing to give up some live-web convenience to reduce the chance that a training model can affect external systems. Whether other labs adopt the same policy will depend on their own workflows and risk assessments.

2. OpenAI and Poland’s NASK will explore AI safety research

OpenAI and Poland’s National Research Institute, NASK, announced on October 23 that they had agreed to explore collaboration on AI safety, security, and technical research. NASK’s AI Safety Research Center also serves as Poland’s AI safety institute. The announcements named possible work on evaluating more capable systems, identifying vulnerabilities, and strengthening safeguards, but did not set out a detailed research timetable or a defined independent-audit scope. OpenAI announcement, NASK announcement

I don’t want to get ahead of the evidence and call this an independent audit already underway. What is confirmed is that a model developer and a public research institute have opened a path to work together. The value will depend on which systems are tested, under what conditions, and how openly the methods and limitations are shared. External research can bring perspectives that an internal team may miss, but a cooperation announcement alone proves neither independence nor safety. The evaluation methods and results will be the more meaningful record.

3. Telecom AI can help restore service, but autonomous networks remain years away

AT&T and T-Mobile told Reuters they are using AI in network operations while strengthening protections for customer data and critical communications infrastructure. Both said fully autonomous networks are still years away. Current deployments focus on bounded tasks such as analyzing trouble tickets or making limited repairs. T-Mobile uses a self-organizing network to adjust nearby antennas when a cell site goes down. The company said that during January’s Winter Storm Fern, coverage returned to about 66% of customers within an hour and 90% within eight hours, with most customers regaining service before power was restored. Reuters report

The practical version of AI I see here is not a machine replacing every network operator. It reads a defined set of conditions, takes a limited action, checks the result, and hands more complex cases to a person. A small configuration error in a telecom network can affect many people, so verifying recovery benefits and failure boundaries matters more than expanding authority. The figures are the company’s account of one storm, not a performance guarantee for every outage. Still, they show that AI’s value can be measured first in narrow jobs that reduce downtime—not in claims of complete autonomy.

4. Twenty-two companies have objected to plans to tokenize their shares

According to a regulatory document reviewed by Reuters, six more companies—including Wells Fargo, Arm Holdings, and Intuit—objected to having their shares tokenized on a trading platform planned by OKX and Intercontinental Exchange (ICE). That brought the total number of objecting companies to 22. The OKX–ICE venture had planned to offer tokens representing 63 stocks. Under the process, issuers must be notified before listing, and a formal objection prevents that company’s shares from being listed there. This is not a ban on every form of stock tokenization; it makes issuer consent a real constraint for this platform’s planned listings. Reuters report

I see the technology for creating a token and the transfer of shareholder rights as separate steps. Twenty-four-hour blockchain trading and faster settlement may be technically possible, but dividends, voting, ownership records, and corporate actions still have to match the underlying shares. Companies may have different reasons for objecting, so I would not treat the list as a blanket rejection of tokenization. The news shows that the next bottleneck is not only writing smart contracts; it is also legal rights and consent from existing market participants.

5. Korea’s third-quarter growth is forecast at 0.7%, with exports more concentrated in chips

A Reuters poll of 16 economists through October 23 put South Korea’s third-quarter real GDP growth at a median 0.7% quarter over quarter, slightly above the 0.6% recorded in the second quarter. That is a survey forecast, not an actual GDP release from the Bank of Korea. The same report said September exports rose 84% year over year to a record $120.9 billion, while semiconductor exports jumped 262.8% to $60.3 billion—nearly half of total exports. Reuters poll and export data

I’m watching the breadth of growth more closely than the forecast alone. Strong chip exports can support the economy and investment, but whether those gains reach household consumption and other industries is a separate question. As semiconductors make up a larger share of exports, shifts in AI capital spending and memory prices can have a larger effect on Korea’s outlook. I won’t quote the GDP forecast as a confirmed result; I’ll revise my view when the actual figure and domestic-demand indicators are available.

6. Brent settled at $116.72 as Middle East supply concerns returned

On October 23, Brent crude futures settled at $116.72 a barrel, up 3.14% for the day; West Texas Intermediate closed at $103.27, up 3.8%. Reuters said Houthi attacks in Saudi Arabia raised concerns about supply disruptions, while authorities in the United Arab Emirates separately reported an oil spill from a tanker. At the time, the cause of the spill and the vessel involved had not been identified. Brent’s settlement was its highest since April 29. Reuters oil report

I don’t see this price move as just another continuation of the same headline. If supply routes remain uncertain, pressure can spread beyond crude to shipping, insurance, electricity, and input costs. An energy-importing economy such as Korea also has to watch the won and inflation. But one daily close cannot establish a lasting shortage or settle the inflation outlook. I would look for evidence of actual supply effects and whether prices remain elevated after the immediate news passes.

What I’m taking away today

Across these stories, the recurring idea is less “autonomy” than “boundaries.” OpenAI has cut live internet access from frontier training; telecom operators let AI help with defined recovery tasks but hold back from fully autonomous networks; and stock tokenization has to pass issuer consent and preserve existing rights.

The economic signals do not point in only one direction. Korea’s exports are strong and concentrated in semiconductors, but its GDP figure is still a poll forecast. Oil has climbed again on supply fears. When I assess AI automation or market optimism, I want to ask not only whether a capability works, but who authorizes it, who can stop it, and who bears the cost when it fails. I’ll watch for actual evaluation results, telecom outage records, issuer consent, Korea’s confirmed GDP, and the path of oil prices.

This is market-structure analysis for information only, not a recommendation to buy or sell any asset. Each reader is responsible for their own decisions.