VASCOM TECHNOLOGY AUTOMATION CO., LTD
Home › News › AI in business: model safety, capability validation, and…

AI in business: model safety, capability validation, and data protection

Three developments show why businesses should validate AI capabilities, assign reviewers, and control data before giving agents tasks.

From building safe models to preventing agents from reading screens, this week’s AI news raises a common question about control. For businesses, that question needs an answer in each specific workflow.

1. Anthropic prioritizes safety amid pressure to grow

VnExpress Số hoá reviews the origins of Anthropic, founded in 1/2021 by Dario and Daniela Amodei and 5 colleagues who left OpenAI. According to the article, the founders wanted to put safety and responsibility at the center of research and products. Anthropic describes its goal through three characteristics: systems that are reliable, interpretable, and steerable. These characteristics directly affect whether results are consistent, how well people can understand a model’s decisions, and how they can guide its behavior.

The article describes two research approaches serving that goal. One seeks ways to observe what happens inside a model as it processes information. The other is Constitutional AI, a method that uses a written set of principles against which the system checks and adjusts its responses. VnExpress also places safety in the context of the company’s need for substantial computing resources, data centers, and staff. As an AI product scales, its safety commitments must continue to be tested in how the product operates.

For businesses, the name of a training method is only a starting point for questions. Your company still needs to define which data a system may access, what actions it may take, and who is responsible for reviewing its results. A model with safety principles does not automatically know each department’s internal rules. Those rules need to become access permissions, approval steps, and procedures for handling answers that are wrong or lack supporting evidence.

2. OpenAI releases a test of AI in drug research

OpenAI introduced MedChemBench on 20/11/2026 to assess models’ reasoning capabilities in medicinal chemistry. The benchmark contains 99 problems drawn from 117 papers and patents, with 761 distinct molecular structures. Models must read structures, interpret experimental evidence, rank compounds, and predict properties. Many problems provide context from a particular point in the research process; later experimental results are held back for comparison with the predictions.

At the xhigh reasoning level, GPT-6.2 Astra scored 84,5% of the available points. GPT-6 Astra scored 83,3%, GPT-6.1 Sol scored 82,0%, and GPT-5.6 Sol scored 58,2%. These percentages are MedChemBench results, not success rates for AI across all tasks. OpenAI also describes a problem in which compound 23 inhibits PRMT5 in cells at a lower concentration, while compound 24 crosses the cell barrier more effectively and breaks down more slowly in a liver enzyme assay. Choosing a compound based on one strong measure can overlook other trade-offs.

The lesson for businesses is to test AI on the types of work they intend to assign it. An aggregate score says little about whether a model will handle your company’s price lists, request forms, or technical records well. An internal test set should include common inputs, cases that are easy to misjudge, and clear scoring criteria. Subject matter experts need to review both answers and their supporting evidence before deciding to use the results in a workflow.

3. Signal restricts screenshots in its mobile app

According to TechCrunch AI on 21/11/2026, Signal is restricting screenshots within its app on Android and iOS. Signal had previously blocked screenshots in its Windows app by default. Signal President Meredith Whittaker raised concerns that AI agents at the operating system level could read information displayed on a screen. Once content has been collected, stored, or sent elsewhere, the messaging app can no longer control how that data is used on its own.

Signal lets users turn off the new restriction and acknowledges that the measure does not address every way an agent might access app data. Blocking screenshots therefore does not mean all sensitive content is protected. This development shows that agent access needs to be considered across the app, device, and operating system. A conversation may be protected within an app but still appear on a screen that another tool is allowed to observe.

For businesses using AI to support daily work, data boundaries need to be defined before enabling screen reading or connecting apps. Your company should know what an agent can see when an employee opens a contract, customer information, or an internal conversation. Rules are needed for permitted tasks, devices used, and situations in which work must stop for human review.

Three steps businesses should take now

  1. List data and access permissions for each AI task. Small and medium-sized businesses in Vietnam can start with document-heavy processes, such as receiving requests or preparing quotations. Give agents access only to the data needed to complete their assigned work.
  2. Test AI on controlled real records before expanding its use. Your company should select routine cases and cases prone to errors, then record the results for subject matter experts to assess. This helps identify tasks that need more data, guidance, or an approval step.
  3. Assign reviewers for outputs that affect customers or operations. A draft, recommendation, or action proposed by an agent needs an appropriate review point before it is sent or carried out. Businesses should specify who can approve it and what information that person must check.

How VASCOM views these developments

VASCOM starts by examining the workflow: where data enters, which steps employees take to process it, and which decisions require an accountable person. From there, a business can consider agents for individual departments, each with specific tasks and access permissions. Testing one workflow allows AI results to be compared with how employees currently handle the work.

VASCOM recommends keeping human reviewers at points that affect customers, money, or sensitive data. Data held by the business needs clear access rules, including when an agent is connected to multiple tools. Once the task scope, data sources, and approvers are defined, a business has a basis for assessing AI through work results and its actual level of control.


Sources

  • VnExpress Số hoá, “Anthropic – from AI laboratory to trillion-dollar empire”, 21/11/2026, vnexpress.net.
  • OpenAI, “Can AI help scientists identify promising drug candidates?”, 20/11/2026, openai.com.
  • TechCrunch AI, “Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance”, 21/11/2026, techcrunch.com.
Technology news Talk to VASCOM
More news

Keep reading

Zalo ☎
VASCOM
Call 0822.577.579 Get a consultation