What Changed in AI This Week? OpenAI Slows Astra, ChatGPT Adds Teen Safeguards, Anthropic Reassesses Risk and Microsoft Reshapes Copilot
What Changed in AI This Week? OpenAI Slows Astra, ChatGPT Adds Teen Safeguards, Anthropic Reassesses Risk and Microsoft Reshapes Copilot
OpenAI slowed Astra training, ChatGPT introduced stronger safeguards for teens, Anthropic reassessed risk, Microsoft reshaped Copilot and AI shopping became a more practical business issue.
It was a quieter week for AI headlines, but the useful developments all point in a similar direction. Capability, security, product packaging, user safety and commercial discoverability are increasingly moving together, which means businesses need to plan around more than model performance alone.
It’s that time of the week again, where we take a look at the most useful AI news stories to have emerged and break them down so you don’t have to spend hours filtering through the noise.
Now, it may in part be down to it being summer, but it was something of a quieter week; nonetheless, it’s been full of important reminders that capability, security, product packaging, user safety and commercial discoverability are now moving together, and businesses need to plan for all five.
Headlines at a glance
OpenAI slows Astra training
A two-week pause in reinforcement learning and a continuing hold on its largest planned run show that frontier capability is now affecting release cadence.
ChatGPT adds stronger teen safeguards
Under-18 accounts are moving into a dedicated experience with learning tools, break reminders and clearer boundaries designed to keep the chatbot from feeling like a human relationship.
Anthropic revises its risk assessment
A previously missed safeguard gap led the company to describe certain catastrophic risks as low but not negligible, rather than very low.
Microsoft consolidates Copilot
The refreshed app brings personal and work sign-ins into one shell, while Group Chat, Podcasts and consumer Deep Research are being retired.
AI shopping becomes an operational issue
Microsoft’s latest guidance puts structured product data, Copilot Checkout and AI visibility measurement at the centre of retail readiness.
Efficient agent models reach AWS
NVIDIA Nemotron 3.5 Lightning is now available through SageMaker JumpStart for high-volume agent tasks where frontier-model cost would be hard to justify.
OpenAI slows Astra as cyber capability starts to affect release cadence
OpenAI said it temporarily slowed the scaling of its latest models after preliminary evaluations suggested that Astra, an upcoming model, may reach its Preparedness Framework’s “Critical” cybersecurity threshold. The company paused reinforcement-learning training for two weeks and says its largest planned frontier run remains on hold while smaller evaluations continue. Read more directly from OpenAI here.
‘A model launch may arrive later’ is hardly news, as delays aren’t exactly uncommon in tech developments. Yet, OpenAI is treating the training infrastructure itself as a security boundary, tightening workload and network isolation, expanding continuous security testing and increasing monitoring across more stages of training. OpenAI estimates that this monitoring can add roughly 20% to the inference compute being watched, although the cost varies by workload.
As models become more capable, safeguards will consume more engineering time and compute, and availability may become less predictable, all of which are useful to know for buyers, especially as the same systems that improve code review and vulnerability discovery can also lower the barrier to attack. These days, security is no longer something vendors can add after the model is finished; it's becoming part of the model-development schedule and economics.
Don't build a critical workflow around an unreleased model or assume every capability jump will arrive on time. Keep a supported fallback model, document which processes depend on a particular vendor, and ask how agent access to code, credentials and the internet is isolated and logged.
Anthropic’s risk report is a governance lesson, not just a safety paper
Anthropic’s August risk report, published at the end of last week, is still shaping this week’s conversation and reaches a similarly sober conclusion. The company now describes the risk of catastrophic harm in one chemical and biological threat category as low but not negligible. It also revised its February assessment from “very low” to “low” after finding that biological safety classifiers had not been applied to most human-feedback vendor traffic for nearly a year. Read Anthropic’s August 2026 Risk Report.
Anthropic says it fixed the issue after learning about it in April, found no evidence of chemical or biological misuse, and that no customer data, internal systems or model weights were exposed. The uncomfortable part is the process lesson, wherein a control can look sensible on paper and still miss a real route through which people or suppliers interact with a system.
For ordinary businesses, though, the parallel is straightforward: if you’ve got an AI policy that covers the approved chatbot but ignores browser extensions, automation tools, contractors, evaluation accounts, and embedded features, it isn't a complete control environment. The route into the model matters just as much as the model name on the invoice.
Map every route by which staff, suppliers and software can send company data to an AI service. Then check whether retention, access controls, logging, and data loss prevention apply consistently across those routes. Governance should follow the data path, not the product logo.
Microsoft consolidates Copilot and retires several familiar features
Microsoft began rolling out an updated Copilot app on 18 August, aiming for one more cohesive experience across web, desktop and mobile. Users can sign in with personal, work, or school accounts and switch between them, while Microsoft says that commercial data boundaries, tenant controls, and compliance protections remain unchanged. Learn more from Microsoft here.
The simplification also comes with some removals, as Group Chat content created by other participants won't carry forward after an account update, Podcasts have been retired, and Deep Research is being retired from the consumer Copilot app. Microsoft 365 Premium subscribers can continue in-depth report work through Researcher, while existing reports remain available through the relevant history or Researcher experience.
It seems that Microsoft is drawing a clearer line between free chat and the paid Microsoft 365 experience, where higher limits, agents and complex multi-step work sit. For small firms that use a mix of personal Microsoft accounts and company-managed Microsoft 365 accounts, the cleaner shell may help. It also makes account discipline more important because a familiar Copilot icon can represent very different data protections depending on how someone signed in.
Check which Copilot account types your team actually uses, not just whether they “use Copilot”. Export any valuable Group Chat media or Podcasts that haven't already migrated, and make work-account use the default for company information. Treat feature retirement as a reminder that AI output should live in governed business systems, not only in chat history.
Microsoft’s AI shopping playbook makes product data a sales issue
Microsoft Advertising used this week to set out a 90-day readiness plan for AI-assisted shopping. Its argument, in short, is that AI assistants make recommendations based on structured facts, so incomplete product feeds can make a retailer effectively invisible. Microsoft points businesses towards an AI-ready feed in Merchant Centre, AI-native ads in Bing and Copilot, Copilot Checkout, a site-based Brand Agent and AI Visibility insights in Microsoft Clarity. Read Microsoft Advertising’s guidance.
Some of that is, unsurprisingly, Microsoft promoting Microsoft’s own advertising stack. But the underlying lesson is bigger than just Bing. Product names, prices, availability, specifications, delivery terms, and return information need to be accurate, structured, and consistent wherever an assistant may retrieve them. Traditional search optimisation focused heavily on earning the click; AI-led discovery also depends on being understandable enough to earn the recommendation.
Retailers should audit their product feed before commissioning a grand “agentic commerce strategy”. Fix missing attributes, stale prices and inconsistent identifiers, then measure which AI services refer traffic and what those visitors do. If you’re a retailer and don’t know where to begin, cleaning your product data is a useful first move.
NVIDIA’s efficient agent model reaches AWS, and the cost message matters
AWS made NVIDIA Nemotron 3.5 Lightning available through SageMaker JumpStart on the 17th of August, a 30-billion-parameter mixture-of-experts model with three billion parameters active at a time, designed for repetitive, high-volume steps in always-on agent workflows. AWS says it can be deployed without configuring the serving infrastructure from scratch and can run on a single supported GPU. Read the AWS announcement.
NVIDIA reports up to 4x higher throughput and up to 30% faster task completion for the workloads it tested; however, these are vendor benchmarks and should be validated against a real process. The message of this story, for you, is to bear in mind that not every step in an agent workflow requires the most expensive frontier model. Planning may require a powerful model, while routine extraction, validation or tool calls may be handled by a smaller specialist.
If you're building agents, measure cost and quality by workflow stage. A routed mix of models can be more economical than sending every request to the same premium model, but only if evaluation, monitoring and fallback behaviour are designed in from the start.
ChatGPT for Teens puts clearer boundaries around how young people use AI
There’s been a lot of talk, and negative stories, about AI and the younger crowd, and frankly, it’s more than a little concerning. That’s why the last story of this week feels like a step in the right direction.
OpenAI began rolling out ChatGPT for Teens on 18th of August, automatically applying the experience when an account is identified as belonging to someone under 18. It brings stronger default safeguards around areas such as self-harm, violence, eating disorders and sexual or graphic content, alongside parental controls and additional safety notifications. This BBC News article dives into more detail, for those curious: OpenAI makes ChatGPT less ‘human’ for teens in new safety update.
One of the more interesting choices is deliberately reducing the sense that the chatbot is a person. OpenAI says teen accounts will receive more frequent break reminders and cues identifying ChatGPT as AI. Its updated rules also say the assistant shouldn’t encourage emotional dependence, imply that it has feelings or consciousness, or use romantic language with under-18s. The experience also adds Study Mode, homework reminders, quizzes and study hours, pushing the product towards guided learning rather than simply handing over answers.
For businesses, schools and parents, this is another sign that AI governance is moving beyond data security and acceptable-use policies, as the same tool can now behave differently depending on the age of the person using it. Similarly, for organisations with apprentices, work-experience students, or younger users, you may need to consider which accounts they provide, which features are appropriate, and where human oversight still matters.
If under-18s can access AI through your organisation, don’t assume the standard adult setup is the right one. Check how age is identified, what protections are enabled by default, whether parental or organisational controls apply, and ensure AI is positioned as a useful tool rather than a substitute for teachers, managers, or real-world support.
The practical takeaway
This week’s thread is that AI is becoming operationally serious. Frontier vendors are slowing work to strengthen controls, while products used by younger people are gaining more deliberate safety boundaries. Product suites are consolidating and retiring features. Retail visibility increasingly depends on machine-readable data, and agent builders are looking for cheaper execution layers rather than using a frontier model for everything.
For SMBs and growing businesses, the sensible response isn’t to chase every announcement, that’s why we don’t do a roundup that’s full of 30 stories a week. Instead, pick a small number of valuable workflows, keep company data inside managed accounts, record where each workflow depends on a vendor or feature, and review the controls around tools, suppliers and integrations. If younger users, apprentices or students can access AI through your organisation, age-appropriate use and safeguarding now belong in that conversation too. Capability still matters, but reliability, governance, safety and cost are what turn it into useful business infrastructure.
Turn AI adoption into a managed business capability
We can help you identify useful workflows, understand the risks around your current tools and put proportionate governance around adoption.