AI Highlights — the whole picture — 2026-09-22 (Tue) Evening News

That morning, two competing AI companies were about to strike a deal that would intentionally drive the other's models. To what extent should safety be determined within the company, and to what extent should it be left to outside parties? The same question was repeated in different forms on loan screening screens, on classroom bulletin boards, and even in stock price movements watched by investors. No single standard has been set yet. The question of who would be in charge of it has been at the forefront of the last half-day.

Mutual verification and creation of measurement standards become practical

Mutual verification and measurement standard creation become a practice

OpenAI and Anthropic are on the verge of agreeing to a deal to subject each other's AI models to severe stress tests, The Information reported. This is an agreement between two competing companies to deliberately drive the other company's models to identify vulnerabilities and malfunctions, and is attracting attention as a move to expose safety verification to outsiders rather than just internal evaluations. Around the same time, OpenAI proposed the creation of a cross-industry global AI standard to guide alignment and RSI (Risk and Safety Infrastructure), according to CNBC, and the company's website also indicated a similar direction under the title ``Building standards for the next stage of AI.'' The way we measure safety itself is beginning to shift from a company's internal rules to a standard shared among companies.

Research that has not yet been peer-reviewed has already begun to give concrete shape to the creation of this scale. A benchmark to measure whether AI assistants working in recruitment, medical care, or finance can follow the rules written in system prompts under pressure. PACT proposed a design that pairs rules to be followed with shortcuts that can be broken quickly and measures the degree of compliance over multiple turns of conversation. The authors report that even the strongest models still have the problem of misapplying rules, and that even the application of normal pressure increases violations. On the other hand, a method called XConf has demonstrated a new way to measure the degree of confidence that determines whether to use an AI answer as is or pass it on to someone else. While existing methods calculate confidence only from the current inference, XConf calls up the records of past models solved and scored, and allows the model to restate its confidence based on past results in similar situations. The authors report that they were able to make the same or better judgments at less cost than the conventional method, which requires solving the problem many times and takes a majority vote.

Furthermore, a mechanism for cooperation has been proposed for those responsible for the research itself. Even if you line up multiple autonomous research agents, if each one starts their search from scratch, they may end up just following the same path. The Agora framework proposed a method for sharing all results, hypotheses, and verification as immutable Git commits, and coordinating a group of agents without a central planning role. While the authors publish records of nearly 12 days of operation, they acknowledge that controlled comparisons are needed to determine whether shared records actually increase discoveries.

Load testing contracts, industry standard recommendations, and benchmarks to measure compliance and confidence all converge toward a single purpose. The question is whose yardstick should we use to measure the extent to which we can trust and entrust AI? The focus is on where the subject holding the measuring stick will expand next.

Rapidly growing AI agents and porous defenses

Meta's personal agent "Muse" surpassed ChatGPT, Grok, and Claude in weekly downloads immediately after its release, CNBC reported. While cross-verification and measurement standards for companies are starting to become a reality, AI agents for consumers are flooding into the market without waiting for standards to be established. ARD #167 (AI: Reset to Zero), which described the U.S.-China AI development competition, including Anthropic's moves, as an unusual situation ``like a cat and a dog living together'' is apt to point out the current situation where multiple camps are on the same playing field, with norms not yet set in stone.

The speed of its spread directly highlights the lag in defense. Amazon has blocked Meta's attempts to gain unauthorized access to its Muse agent, Currently.com reported. Agents that act autonomously on the web on behalf of users create new targets for attackers because they carry credentials and session privileges with them. The Hacker News weekly roundup juxtaposes the Cisco zero-day vulnerability, remote code execution (RCE) targeting AI agents, the proliferation of ClickFix-style attacks, and browser hijacking in the same week, showing a parallel spread of attack methods rather than isolated incidents.

The number of downloads, which is an indicator of penetration, and the indicator of protection, which is blocked unauthorized access and RCE, move at completely different speeds on the same time axis. The reality is that millions of people are entrusting personal information and operational authority to agents, and the mechanisms to protect it are only reactive. This asymmetry means that the verification framework that has been built up between companies has not yet reached consumer products.

This discrepancy in the speed of adoption and defense posture is pushing the question of who decides safety standards from internal corporate discussions to the everyday lives of ordinary users.

The economic effects of AI have begun to be expressed in numbers

Numbers measuring the economic impact of AI have skyrocketed in recent weeks. Axios reports that AI could become a factor in inflation. Fox Business reported that U.S. Federal Reserve Board member Neel Kashkari also expressed caution that the expansion of AI-related investment could push up upward pressure on prices. The same article also introduces ``Air 7'', which is not ``Magnificent 7'', an AI-related stock chosen by Kevin Mahn, and points out that the Fed's interest rate decisions influence the price movements of this group of stocks. Benzinga similarly published an article warning ETF investors to be wary of AI-driven inflationary pressures, citing the State Street Utilities Select Sector SPDR ETF as an example.

Semiconductor price trends have also begun to be discussed in concrete numbers. According to 24/7 Wall St., AMD stock rose 5% following a report that chip prices rose 10%, and the partnership between NVIDIA and Taiwan Semiconductor (TSMC) also received renewed attention. Additionally, TradingView reports that NVIDIA's CEO has made the most in-depth comments ever regarding data centers for AI. Both data center demand and semiconductor prices appear to support the spread of AI investment into the real economy.

There are also estimates of the impact on the local economy. According to a report from the Inter-American Development Bank (IDB) reported by IndexBox, the spread of AI could boost Latin America's GDP by up to 5.1%, while reducing wage risk by 20.9%. The effects of AI are presented as positive numbers in terms of both growth and employment. The Business Journals reports that as the use of AI expands, new liability concerns are emerging for companies, indicating that behind the economic benefits, governance challenges are also growing at the same time.

On the other hand, IBM's research points to changes in human resources behind the numbers. According to Virtualization Review, an IBM study found that AI could erode the very skills that employers need most. The growth numbers expressed in terms of GDP, stock prices, and chip prices, and the actual situation on the ground where needed human skills are being lost, are two faces of the same expansion of AI that are progressing simultaneously. This duality leads to the next question of what standards should be used to measure what.

Banks increase investment in AI and increase vigilance at the same time

In the field of bank lending, AI is beginning to enter the early stages of credit decisions. An analysis by the Federal Reserve Bank of San Francisco points out that the introduction of AI is having an impact on the efficiency of bank loan review processes and credit decisions, showing that lending practices themselves are becoming a testing ground for the use of AI (Federal Reserve Bank of San Francisco). Bank of America was initially skeptical about investing in AI startups, but is reportedly changing its tune and moving forward with funding AI ventures (PYMNTS.com). The move by major banks to proceed with investment after careful verification proves that the use of AI is not a passing fad, but is becoming incorporated into the structure of lending operations.

Similar movements are accelerating in the mortgage business field. At the "Digital Mortgage 2026" showcase, 20 AI solutions for lenders were introduced, and companies competed to implement AI tools in each process of screening, document processing, and customer service (National Mortgage News). From loan underwriting to mortgage processing, AI adoption is moving beyond individual pilots to becoming a standardized option across industries.

On the other hand, the same technological foundation is also emerging as a risk aspect. Experts are calling for vigilance as text message scams posing as banks and AI-based scams are on the rise (wafb.com). AI-generated texts and sounds are becoming more natural, making it harder for customers to distinguish between fake messages and legitimate bank notifications. This situation in which attacks and defenses are progressing on both sides of the same technology shows that banks are being forced to make investment decisions and security measures at the same time. The Business Journals reports that banks are adopting unique approaches to AI rather than uniformly, highlighting the fact that the selection of investment targets and the delineation of the scope of implementation are left to each bank's management judgment (The Business Journals).

Banks are simultaneously increasing the speed at which they incorporate AI and their vigilance in three areas: loan appraisal, mortgage processing, and fraud prevention. The outcome of this tug-of-war is also beginning to be questioned in fields outside of finance, such as education and academic research.

Tug of war over how to use AI in educational settings

A tug-of-war over how to use AI in educational settings has surfaced, both in terms of subsidies and field regulations. Hudson County Community College is moving to support faculty innovation through the AI Spark grant, taking a position to institutionally support faculty in incorporating AI into their lesson design and work (TAPinto). Meanwhile, New York City public schools have announced a policy banning the use of AI in middle schools starting this year (ABC News). Within the same educational administration, there is a structure in which measures to encourage utilization and measures to restrict use are running at the same time.

Concerns from the field are more immediate. The head of a teachers' union says AI will harm children's critical thinking skills (The Hill, NewsNation). An essay on educational settings in Latin America also argues that while AI may provide students with the answers themselves, the role of teaching them judgment should remain as the role of humans (UPI). The article published in The Good Men Project discusses the importance of teaching students what AI cannot do, and takes the position that the limitations of AI as a tool should be incorporated into educational content. Universities are expanding the use of AI through subsidies, public school districts are restricting its use, and teachers' unions are concerned about the decline in critical thinking skills.All three parties have different answers to the same question: ``How do we incorporate AI into education?''

Behind this tug-of-war is a broader perception of risk over the extent to which AI output can be accepted unconditionally. Omdia points out that the risks of AI without human involvement are increasing, and this is a concern that is common to all situations where AI is used, not just in educational settings. The core of the teachers' union's argument is that if students get into the habit of accepting AI's answers as they are, the core of education, which is the development of critical thinking skills, will be affected.

Neither those providing subsidies nor those restricting its use have reached the conclusion that AI itself should be eliminated. Rather, we are at the stage where we are exploring from each perspective how much human judgment should be involved. The question of who should take the lead in drawing this line is beginning to be asked not only in education, but also in a broader context, such as the extent to which AI output can be trusted and incorporated into work and research.

Mutual verification agreements, proposals for the creation of industry standards, banks' credit decisions, subsidies and bans in educational settings all all have their roots in a single movement. Multiple workplaces are simultaneously faced with a situation where someone has to decide how much to trust AI. It has not yet been determined in any section who will have the deciding factor. Is it a company's internal regulations, a cross-industry standard, the discretion of the workplace, or the discretion of the regulatory authority? Without deciding on the answer, only judgment continues to move forward.

Q1: When incorporating AI into the work at hand, who do you think can ultimately verify the correctness of the decision? Q2: To what extent are you willing to delegate the standards for determining the scope of AI use to someone outside your organization? Q3: To what extent can you justify the decision to proceed with investment and introduction before standards are established?

Finally, three questions

- When incorporating AI into the work at hand, who do you think can ultimately verify the correctness of the decision? Q2: To what extent are you willing to delegate the standards for determining the scope of AI use to someone outside your organization? Q3: To what extent can you justify the decision to proceed with investment and introduction before standards are established? - To what extent are you willing to delegate the standards that define the scope of AI use to someone outside your organization? Q3: To what extent can you justify the decision to proceed with investment and introduction before standards are established? - To what extent can we justify the decision to proceed with investment and implementation before standards have been established?