October 22

Your Daily Note,

Walmart is preparing for AI agents to shop on customers’ behalf, OpenAI is bringing its models to Google Cloud, and researchers have used AI to uncover serious vulnerabilities in PHP, a programming language used by millions of websites.

AI is moving from answering questions to taking action. It can help people shop, write software and analyse business data. But as these systems become more capable, new questions emerge: who is responsible when an agent makes a mistake, how do companies keep their systems secure, and does using AI actually make work easier?

In today’s Noted:

AI uncovers decades-old PHP vulnerabilities. Millions of websites may need attention.

We noted:

Researchers using AI-assisted security tools have uncovered two critical vulnerabilities in PHP, a programming language that powers a large share of the web. The US National Security Agency issued a warning urging organisations to take the flaws seriously. Patches have been released.

The discovery highlights both sides of AI-assisted security: AI can help defenders find hidden weaknesses, but similar capabilities could also help attackers exploit them.

Why it matters

Website owners and developers using affected PHP versions should check the official PHP security advisories and upgrade to a supported, patched release. PHP 8.1 and older versions are no longer supported, so simply applying an old patch may not be enough for every installation. If you run a WordPress site, this is a server-level PHP issue, not a vulnerability exclusive to WordPress. Contact your hosting provider if you are unsure which PHP version your site uses.

AI-assisted security research could make it faster and cheaper to discover vulnerabilities that have survived for years. That creates an advantage for defenders, but also puts pressure on organisations to patch systems before attackers adapt the same techniques.

Sneaky useful: security is not only about preventing new attacks. Keeping old software supported and patched can be just as important as investing in the latest security tools.

Walmart wants AI agents to shop for customers. Amazon is blocking them.

We noted:

Walmart has introduced an Agent Gateway designed to let approved AI agents search its online inventory and make purchases on behalf of users. The move contrasts with Amazon, which has blocked third-party AI agents from shopping on its website.

Walmart’s Agent Gateway uses the Model Context Protocol, or MCP, to connect approved AI agents to its shopping systems. An agent can search products and perform transactions rather than merely recommend what someone should buy.

Why it matters

Shopping could increasingly start with a request such as finding groceries for the week within a budget, comparing products and preparing an order. The experience will depend on which retailers allow access, what permissions users grant and whether the agent can handle payment, substitutions and checkout correctly.

Retailers are deciding whether AI agents should be treated as customers, partners or unauthorised automated traffic. That decision could shape which products agents can discover, which stores they recommend and where consumers ultimately spend money. If you sell products online, start thinking about whether your product information is easy for AI systems to understand. Accurate descriptions, prices, stock availability and clear purchasing pathways may become increasingly important as shopping agents grow.

OpenAI models are coming to Google Cloud. Businesses get another way to access them.

We noted:

OpenAI and Google Cloud have announced a partnership to make OpenAI’s frontier models and Codex coding workflows available through Google Cloud’s Model Garden. This gives enterprise customers another way to access OpenAI technology within their existing cloud environment.

Model Garden is Google Cloud’s platform for discovering and deploying AI models. Through the new partnership, customers will be able to use OpenAI models alongside Google’s infrastructure, access controls and data-governance systems.

Why it matters

Businesses already operating on Google Cloud may eventually be able to evaluate and deploy OpenAI models without building a separate infrastructure arrangement from scratch. This does not mean the models are already available to every Google Cloud customer, so organisations will still need to assess access, pricing, security requirements and suitability for their workloads.

AI competition is no longer just about which company builds the best model. It is also about which company makes that model easiest to access where customers already work. When choosing an AI model for your product, compare the full deployment cost, not just the model’s API price. Existing cloud commitments, security requirements and integration work can all influence the final bill.

AI promises to save time. Managers say it is creating more work.

We noted:

A new survey of 1,700 directors and senior managers across seven countries, commissioned by Monday.com and conducted by Nielsen, found that AI adoption is creating new management challenges even as enthusiasm remains high. More than half of surveyed senior leaders said evaluating employees’ work had become harder after AI adoption.

AI can help employees produce polished reports, code and presentations faster. But managers may struggle to determine how much work was done by employees, how much was generated by AI, and whether the output is actually correct.

Why it matters

These are survey findings, not proof that every company is experiencing the same results. Businesses adopting AI may need to redesign their workflows, clarify who is responsible for AI-assisted work and establish better review processes. Managers may increasingly evaluate how people use AI, verify its output and make decisions, rather than focusing only on how much they produce. Simply giving employees AI subscriptions is unlikely to solve every productivity problem.

Productivity gains can be undermined if the time saved on drafting is spent checking errors, managing additional tools and coordinating automated work. AI adoption is becoming an organisational challenge as much as a technology decision. Before introducing another AI tool, identify the specific bottleneck it should remove. Then measure the entire workflow, including review and correction time, instead of counting how many tasks AI completes.

Meta says Muse has over 10 million users. Shopping is becoming a major use case.

We noted:

Meta said during its third-quarter earnings call that its AI agent app Muse has attracted more than 10 million users. CEO Mark Zuckerberg also highlighted shopping as a major activity, with some users making purchases through the agent.

These usage and spending figures are reported by Meta and should not be confused with independently verified long-term performance.

Why it matters

AI assistants may increasingly help people discover products, compare options and complete purchases. If these experiences become more convenient, consumers may spend less time navigating individual retailer websites. However, an AI agent’s recommendation is not a guarantee that a product is suitable, that a price is the best available or that a transaction will go smoothly. Users should still review purchases, permissions and payment details.

Meta’s figures suggest that AI agents may be moving beyond casual conversation towards commercial activity. If people begin trusting assistants to help them shop, the competition for consumer attention could shift from websites and apps towards the agents that guide purchasing decisions. For online businesses, ask whether your products can be found and understood outside your own website.

Products worth trying

 

Everything else in AI

OpenAI wants the same safety standards wherever its models are used

OpenAI says it is aligning safety requirements across services that provide access to its models, including hosting partners, resellers and model-routing platforms. Its updated Usage Policies are scheduled to take effect on 20 November 2026 and will require services offering direct access to its models, such as model routers and AI gateways, to obtain OpenAI’s review and approval.

The company says partners must maintain safeguards, investigate potential misuse and coordinate enforcement. The goal is to prevent restrictions from being bypassed simply by accessing a model through another provider. Developers using third-party gateways should review the updated requirements before the effective date.

 

Blackstone says fears about AI infrastructure debt are overblown

Blackstone is investing across AI infrastructure, from data centres and computing capacity to AI models and applications. At TechCrunch Disrupt, Blackstone’s Jas Khaira argued that concerns about debt financing for the AI buildout are overstated, pointing to the firm’s ability to provide different forms of capital.

The wider issue is that AI infrastructure requires far more upfront investment than many software businesses. Whether those investments pay off depends on future demand, utilisation and the revenue generated by the infrastructure. Khaira’s view is an investor’s perspective, not proof that the debt risks are settled.

 

Recursive launches Flint, a coding model that can run on a laptop

AI startup Recursive has released Flint, a free coding model designed to run locally on a laptop. The company says its automated research system, the Eureka Machine, helped build much of the model in under four months.

Flint is an early example of using AI to assist in the development of another AI model. Because it can run on a user’s own computer, it can also help developers work offline, keep code local and avoid usage-based cloud fees. The model is available on Hugging Face under the Apache 2.0 licence. The company’s claims about how it was built should be distinguished from independent evidence of its performance.

 

NVIDIA introduces Kumo Relational for predictions from business databases

NVIDIA’s Kumo Relational is a pretrained model designed to make predictions directly from structured enterprise databases. It can use historical examples to estimate outcomes such as customer churn, demand and fraud without requiring task-specific training or manually engineered features for each prediction.

NVIDIA reports results across 41 tasks and 15 databases, including performance gains against selected machine-learning benchmarks. Those are vendor-reported evaluations, not a guarantee that the model will outperform existing systems on every business dataset. The potential benefit is reducing the work needed to turn connected business data into useful predictions.

 

VocabX claims a 77% inference-cost reduction for a specialised model

Dharma AI says its VocabX method reduced measured inference costs by 77% in a production essay-grading system by specialising the model’s vocabulary for its task. The work changes how text is tokenised, reducing the number of tokens needed for common outputs and the size of the vocabulary the model must score.

The result is based on a specific Portuguese essay-grading use case, and the accompanying research paper is still forthcoming. It is an interesting reminder that efficiency gains do not always require a larger model. For teams running a narrow, repetitive AI task at scale, the way a model processes text may also be worth optimising.

 

AT&T is using OpenAI to improve its in-house legal workflows

AT&T is working with OpenAI to extend LegalEdge, its initiative for applying AI to legal work. The planned system aims to help legal professionals gather relevant documents, locate past decisions, classify incoming requests and identify the context needed to respond.

The company says the aim is to help lawyers spend less time collecting information and more time applying their legal judgement. The important boundary is that AI is intended to support the process, not replace the lawyers responsible for decisions.

 

IBM updates Granite Speech for transcription, translation and speaker attribution

IBM has introduced Granite Speech 5.0 2B and updated two models in its Granite Speech collection. The models target different needs, including multilingual transcription and translation, faster transcription on local devices, and identifying who spoke and when in recordings.

The collection is released under the Apache 2.0 licence. For businesses handling meetings, customer calls or audio records, the update offers more options for building speech workflows without relying entirely on a hosted proprietary service. Performance will still depend on recording quality, language and the task being performed.

 

ChatGPT Ads expands to 26 more countries

OpenAI says ChatGPT Ads is expanding to 26 additional countries, bringing availability to 89 countries worldwide. The company also reports that more than 100,000 advertisers have run campaigns since the programme launched.

Ads appear on Free and Go plans, while Plus, Pro and Enterprise subscriptions remain ad-free, according to OpenAI. The expansion gives more businesses access to advertising placements during product discovery and comparison conversations. OpenAI says ads are labelled and separate from its answers, and that it does not sell customer data to advertisers. Advertiser access and availability vary by market.

 

Jensen Huang’s AI duck stole the show at NVIDIA GTC

A tiny robot duck became one of the more memorable moments of NVIDIA CEO Jensen Huang’s GTC Berlin keynote on October 21. Four Microducks, developed by Pollen Robotics, took the stage for a live demonstration. They were meant to sit, stand, take a nap and even perform somersaults. The funny part? They did not always follow instructions. One kept moving when Huang told the group to nap, while another needed a reminder to pay attention.

But Microduck is more than a cute robot. Standing just 25 cm tall, it is an open-source robotics platform designed to learn new behaviours. NVIDIA used the demonstration to show how AI agents, simulation tools and computing hardware can work together to train a robot before deploying its learned skills to the physical machine.

See you tomorrow
same time, different note.

~ Stackari

 

P.S. Found something interesting today? Try it, test it, see what you can make. More News and Useful Guides in the making!