VASCOM TECHNOLOGY AUTOMATION CO., LTD
Home › News › AI Strains Cyber Defenses, OpenAI Tightens Controls, Agents…

AI Strains Cyber Defenses, OpenAI Tightens Controls, Agents Affect Commerce

Three developments on 18/11 show businesses need to review legacy systems, control AI permissions and prepare sales channels for agents.

On 18/11/2026, three news stories raised questions about cybersecurity, AI deployment and the future of online sales channels. For small and medium-sized businesses, the common need is to know which systems are connected, who has permission to act and who is responsible when risks arise.

1. AI shortens the time needed to exploit security vulnerabilities

Speaking with VnExpress Số hoá, Beenu Arora, co-founder and CEO of Cyble, said AI is helping attackers investigate vulnerabilities and create exploit code faster. He cited PHP2Shell as an example: related flaws had existed in PHP for decades, but turning them into working exploit code previously required many complex steps. According to the interview, researchers used an AI model to discover the vulnerability in late August and disclosed it on 21/10.

Arora said AI could reduce the time needed to prepare and launch an attack from several months to several days. Defenders, meanwhile, still have to identify affected systems, responsible staff, schedules for downtime to apply fixes and internal approval steps. Cyble Research and Intelligence Labs recorded 24 ransomware attacks against Vietnamese organizations from January to September 2026, compared with 19 in all of 2025. The reporting periods differ, but the figures show that this risk warrants attention.

For businesses, a system that has run reliably for many years still needs security checks. Admin pages, test environments, rarely used websites or old servers connected to the Internet may fall outside routine monitoring. Businesses should know who manages each system and under what conditions it can be updated. When a vulnerability is found, that information helps the responsible team respond faster.

2. OpenAI outlines risk controls from training to deployment

On 18/11/2026, OpenAI described its process for assessing risk at each stage of AI model development. The company said it may pause internal training or hold back a model such as GPT‑6.1 Astra when more time is needed to meet safety standards. OpenAI has also added safety cases to the process and requires more evidence of appropriate behavior throughout training.

According to the announcement, OpenAI organizes safeguards into four layers and reviews them across three stages: research, internal deployment and external deployment. In the research stage, the company emphasizes isolated environments, virtual machines and disconnecting direct Internet access for advanced model training. As a model sees wider use, monitoring and controls at the system level become more important. OpenAI said it assesses both the model and its safeguards before moving to the next stage.

This does not mean a business must follow the same process OpenAI uses to train models. Still, businesses can ask similar questions when bringing AI into their work: what data can the tool see, what tools can it use and who must approve an action before it has a real-world effect? With limited resources, a business can start with one specific process, document access permissions and check the results before expanding.

3. AI agents could change the path from search to purchase

According to TechCrunch on 18/11/2026, Microsoft CEO Satya Nadella predicted that AI agents would change the economics of the web, especially models that depend on advertising and visits from people. He said that when agents search for information or complete transactions on users’ behalf, the way platforms attract customers and earn revenue from traffic could change. This is a prediction about the direction of development, not a conclusion that all current sales channels have already changed.

The article gives a specific example: Amazon is blocking AI agents from accessing its retail website to browse products and make purchases on users’ behalf. Nadella suggested that another platform, such as Shopify, might allow agents access and gain an advantage as a result. The example shows that allowing agents to access product catalogs could become a business decision alongside familiar decisions about advertising and website design.

For small and medium-sized businesses selling online, the immediate task is to keep product information, prices, stock status and transaction terms clear and consistent across channels. Businesses also need to decide which tasks agents can support and which steps still require human confirmation. As the purchasing journey changes, missing or conflicting data will make it harder for both customers and intermediary tools to make accurate decisions.

Three steps businesses should take now

  1. List systems exposed to the Internet. Include admin pages, test environments and rarely used websites, and identify who is responsible for each system. For small and medium-sized businesses, this list is a starting point for prioritizing vulnerability checks and fixes.
  2. Limit each AI application’s permissions to its task. Define what data it can read, what actions it can take and when it must wait for human approval. Businesses should test it in a narrowly defined process, record errors and adjust permissions before assigning agents more work.
  3. Review data used for online sales. Check product information and order confirmation rules across the channels the business uses. Clear data helps employees work consistently and provides a basis for assessing how agents might support customers later.

VASCOM’s view of these developments

VASCOM believes the first step is to examine how a business actually works: where data flows, which systems are connected and which decisions require a person to take responsibility. Based on that review, a business can choose one clearly defined task for an AI trial while deciding how to check the results and respond when the tool provides incorrect information.

When expanding, VASCOM recommends designing agents for specific departments and tasks, with corresponding access permissions. Actions affecting transactions, important data or operating systems should require human approval. Where business data is stored and the rules for accessing it should be defined as part of process design, so AI use remains within the business’s ability to manage it.


References

  • VnExpress Số hoá, “AI is widening the gap between cyberattack and defense”, 18/11/2026, vnexpress.net.
  • OpenAI, “Managing frontier AI risk: From training to deployment”, 18/11/2026, openai.com.
  • TechCrunch AI, “Satya Nadella predicts AI agents will reshuffle the economics of the web”, 18/11/2026, techcrunch.com.
Technology news Talk to VASCOM
More news

Keep reading

Zalo ☎
VASCOM
Call 0822.577.579 Get a consultation