Microsoft has confirmed it is adding Anthropic’s Claude models to its Copilot AI assistant, giving enterprise users a new option alongside OpenAI for handling complex tasks in Microsoft 365.

Microsoft Expands Model Choice In Copilot

Microsoft has begun rolling out support for Claude Sonnet 4 and Claude Opus 4.1, two of Anthropic’s large language models, within Copilot features in Word, Excel, Outlook and other Microsoft 365 apps. The update applies to both the Copilot “Researcher” agent, used for generating reports and conducting deep analysis, and Copilot Studio, the tool businesses use to build their own AI assistants.

The move significantly expands Microsoft’s model options. Until now, Copilot was powered primarily by OpenAI’s models, such as GPT‑4 and GPT‑4 Turbo, which run on Microsoft’s Azure cloud. With the addition of Claude, Microsoft is now allowing businesses to choose which AI model they want to power specific tasks, with the aim of offering more flexibility and improved performance in different enterprise contexts.

Researcher users can now toggle between OpenAI and Anthropic models once enabled by an administrator. Claude Opus 4.1 is geared towards deep reasoning, coding and multi‑step problem solving, while Claude Sonnet 4 is optimised for content generation, large‑scale data tasks and routine enterprise queries.

Why Microsoft Is Doing This Now

Microsoft has said the goal is to give customers access to “the best AI innovation from across the industry” and to tailor Copilot more closely to different work needs. However, the timing also reflects a broader shift in Microsoft’s AI strategy.

While Microsoft remains OpenAI’s largest financial backer and primary cloud host, the company is actively reducing its dependence on a single partner. It is building its own in‑house model, MAI‑1, and has recently confirmed plans to integrate AI models from other firms such as Meta, xAI, and DeepSeek. Anthropic’s Claude is the first of these to be made available within Microsoft 365 Copilot.

This change also follows a wave of high‑value partnerships between OpenAI and other tech companies. For example, in recent weeks, OpenAI has secured billions in new infrastructure support from Nvidia, Oracle and Broadcom, suggesting a broader distribution of influence across the AI landscape. Microsoft’s latest move helps hedge against any future change in the balance of that relationship.

Microsoft And Its Customers

The introduction of Claude into Copilot is being made available first to commercial users who are enrolled in Microsoft’s Frontier programme, i.e. the early access rollout for experimental Copilot features. Admins must opt in and approve access through the Microsoft 365 admin centre before staff can begin using Anthropic’s models.

Importantly, the Claude models will not run on Microsoft infrastructure. Anthropic’s AI systems are currently hosted on Amazon Web Services (AWS), meaning that any data processed by Claude will be handled outside Microsoft’s own cloud. Microsoft has made clear that this data flow is subject to Anthropic’s terms and conditions.

This external hosting has raised concerns in some quarters, particularly for organisations operating under strict compliance or data residency requirements. Microsoft has responded by emphasising the opt‑in nature of the integration and the ability for administrators to fully control which models are available to users.

For Microsoft, the move appears to strengthen its claim to be a platform‑agnostic AI provider. By integrating Anthropic alongside OpenAI and offering seamless switching between models in both Researcher and Copilot Studio, Microsoft positions itself as a central point of access for enterprise AI, regardless of where the models originate.

Business Relevance And Industry Impact

The change is likely to be welcomed by business users seeking more powerful or specialised models for specific workflows. It may also create new pressure on OpenAI to continue improving performance and pricing for enterprise use.

From a competitive standpoint, Microsoft’s ability to offer Claude inside its productivity suite puts further distance between Copilot and rival AI products from Google Workspace and Apple’s AI integrations. It also allows Microsoft to keep pace with fast‑moving developments in multi‑model orchestration, the ability to run different tasks through different models depending on context or output goals.

For Microsoft’s competitors in the cloud and productivity space, the integration also highlights a growing interoperability challenge. Anthropic is mainly backed by Amazon, and its models run on both AWS and Google Cloud. Microsoft’s decision to incorporate those models into 365 tools represents a break from traditional cloud loyalty and suggests that, in the era of generative AI, usability and capability may matter more than where the models are hosted.

The Google Data Commons Update

While Microsoft is focusing on model integration, Google has taken a different step by making structured real‑world data easier for AI developers to use. This month, it launched the Data Commons Model Context Protocol (MCP) Server, a new tool that allows developers and AI agents to access public datasets using plain natural language.

The MCP Server acts as a bridge between AI systems and the vast Data Commons database, which includes datasets from governments, international organisations, and local authorities. This means that developers can now build agents that access census data, climate statistics or economic indicators simply by asking for them in natural language, without needing to write complex code or API queries.

The launch aims to address the two long‑standing challenges in AI of hallucination and poor data quality. For example, many generative models are trained on unverified web data, which makes them prone to guessing when they lack information. Google’s approach should, therefore, help ground AI responses in verifiable, structured public datasets, improving both reliability and relevance.

ONE Data Agent

One of the first use cases is the ONE Data Agent, created in partnership with the ONE Campaign to support development goals in Africa. The agent uses the MCP Server to surface health and economic data for use in policy and advocacy work. However, Google has confirmed that the server is open to all developers, and has released tools and sample code to help others build similar agents using any large language model.

For Google, this expands its role in the AI ecosystem beyond model development and into data infrastructure. For developers, it lowers the technical barrier to creating trustworthy data‑driven AI agents and opens up new opportunities in sectors such as education, healthcare, environmental analysis and finance.

What Does This Mean For Your Business?

The addition of Claude to Microsoft 365 Copilot marks a clear move towards greater AI optionality, but it also introduces new complexities for both Microsoft and its enterprise customers. While the ability to switch between models gives businesses more control and the potential for improved task performance, it also means IT teams must assess where and how their data is being processed, especially when it leaves the Microsoft cloud. For some UK businesses operating in regulated sectors, this could raise concerns around data governance, third-party hosting, and contractual clarity. Admin-level opt-in gives organisations some control, but the responsibility for managing risk now falls more squarely on IT decision-makers.

For Microsoft, this is both a technical and strategic milestone. The company is reinforcing its Copilot brand as a neutral gateway to the best models available, regardless of origin. It sends a signal that AI delivery will be less about vendor exclusivity and more about task-specific effectiveness. For competitors, the integration of Anthropic models into Microsoft 365 may accelerate demand for open, composable AI stacks that can handle model switching, multi-agent coordination, and fine-grained prompt routing, especially in workplace applications.

Google’s decision to open up real-world data through the MCP Server supports a different but equally important part of the AI ecosystem. For example, many UK developers struggle to ground their AI agents in reliable facts without investing heavily in custom pipelines. The MCP Server simplifies this process, making structured public data directly accessible in plain language. If adopted widely, it could help reduce hallucinations and increase the usefulness of AI across sectors such as policy, healthcare, sustainability, and finance.

Together, these announcements suggest that the next phase of AI will be shaped not only by which models are most powerful, but also by who can offer the most useful data, the clearest integration paths, and the most practical tools for real-world business use. For UK organisations already exploring generative AI, both moves offer new possibilities, but also demand closer scrutiny of how choices around models and data infrastructure will affect operational control, user trust, and long-term value.

Cybercriminals are increasingly targeting employees as a way into company systems, with insider threats now posing a serious and growing risk.

In one recent case, a BBC reporter revealed how a ransomware gang tried to recruit him through a messaging app, offering a share of a ransom if he provided access to BBC systems. The attempt escalated into an MFA bombing attack on his phone, a method used to pressure targets into approving login requests.

This form of insider targeting is becoming more common. For example, the UK’s Information Commissioner’s Office recently found that over half of insider cyber attacks in schools were carried out by students, often using guessed or stolen credentials. In the private sector, insiders have caused major breaches, including a former FinWise employee who accessed data on nearly 700,000 customers after leaving the firm.

Security researchers warn that ransomware groups now actively seek staff willing to trade access for money, rather than relying solely on technical exploits.

To reduce the risk, businesses are advised to enforce strong offboarding, monitor user behaviour, implement phishing-resistant MFA, and raise staff awareness about insider recruitment tactics.

A research team in Denmark is building an AI‑driven robot to refurbish laptops at scale, offering a practical route to reduce e‑waste while creating new value for businesses.

RoboSAPIENS

At the Danish Technological Institute (DTI) in Odense, robotics researchers are developing a system that uses computer vision, machine learning and a robotic arm to automate common refurbishment tasks on used laptops. The project is part of RoboSAPIENS, an EU‑funded research initiative coordinated by Aarhus University that focuses on safe human‑robot collaboration and adaptation to unpredictable scenarios.

DTI’s contribution to the programme centres on robot-assisted remanufacturing. The goal is to design systems that can adapt to product variation, learn new disassembly processes, and maintain high safety standards even when faced with unfamiliar conditions. DTI’s Odense facility hosts dedicated robot halls and test cells where real‑world use cases like this are trialled.

What The Robot Can Do And How It Works

The DTI prototype has been trained to carry out laptop screen replacements, a time‑consuming and repetitive task that requires precision but often suffers from low labour availability. The system does this by using a camera to identify the laptop model and selects the correct tool from a predefined set. It then follows a sequence of learned movements to remove bezels, undo fixings, and lift out damaged screens for replacement.

The robot currently handles two laptop models and their submodels, with more being added as the AI’s training expands. Crucially, the system is designed with humans in the loop. For example, if it encounters unexpected variables, such as an adhesive where it expects a clip, or a screw type it hasn’t seen, it alerts a technician for manual intervention. This mixed‑mode setup allows for consistent output while managing the complexity of real‑world devices.

The Size And Urgency Of The E‑Waste Problem

Electronic waste / E‑waste is the fastest‑growing waste stream in the world. E-waste typically refers to items like discarded smartphones, laptops, tablets, printers, monitors, TVs, cables, chargers, and other electrical or electronic devices that are no longer wanted or functioning. The UN’s 2024 Global E‑Waste Monitor reports that 62 million tonnes of electronic waste were generated globally in 2022, with less than 25 per cent formally collected and recycled. If current trends continue, global e‑waste is expected to reach 82 million tonnes by 2030. That is roughly equivalent to 1.5 million 40‑tonne trucks, enough to circle the Earth.

Unfortunately, the UK is among the highest generators of e‑waste per capita in Europe. Although progress has been made under the WEEE (Waste Electrical and Electronic Equipment) directive, much of the country’s used electronics still go uncollected, unrepaired or end up being recycled in ways that fail to recover valuable materials.

The Benefits

For IT refurbishment firms and IT asset disposition (ITAD) providers, robotic assistance could offer some clear productivity gains. Automating standard tasks such as screen replacements could reduce handling time and increase throughput, while also reducing strain on skilled technicians who can instead focus on more complex repairs or quality assurance.

Mikkel Labori Olsen from DTI points out that a refurbished laptop can actually sell for around €200, while the raw materials reclaimed through basic recycling may only be worth €10. As Olsen explains: “By changing a few simple components, you can make a lot of value from it instead of just selling the recycled components”.

Corporate IT buyers also stand to benefit. For example, the availability of affordable, high‑quality refurbished laptops reduces procurement costs and supports carbon reporting by lowering embodied emissions compared to buying new equipment. For local authorities and public sector buyers, refurbished devices can also be a practical tool in digital inclusion schemes.

Manufacturers may also see long‑term benefits. As regulation around ‘right to repair’ and product lifecycle responsibility tightens, collaborating with refurbishment programmes could help original manufacturers retain brand control, limit counterfeiting, and benefit from downstream product traceability.

Challenges Technical Barriers

Despite its promise, robotics in refurbishment faces multiple challenges and barriers. For example, one of the biggest is product variation. Devices differ widely by brand, model, year and condition. Small differences in screw placement, adhesives, or plastic housing can trip up automation systems. Expanding the robot’s training set and adaptability takes time and requires high‑quality datasets and machine learning frameworks capable of generalisation.

Device design itself is another barrier. For example, many modern laptops are built with glued‑in components or fused assemblies that make disassembly difficult for humans and robots alike. While new EU rules will require smartphones and tablets to include removable batteries by 2027, current generation devices often remain repair‑hostile.

Safety is also critical. Damaged batteries in e‑waste can pose serious fire risks. Any industrial robot working with used electronics must be designed to detect faults and stop operations immediately when hazards are detected. The DTI system integrates vision and force sensors and follows strict safety protocols to ensure safe operation in shared workspaces.

Cost also remains a factor. For example, integrating robotic systems into refurbishment lines requires upfront investment. Firms will, therefore, need a steady supply of similar product types to ensure return on investment. For this reason, early adopters are likely to be larger ITAD providers or logistics firms working with bulk decommissioned equipment.

Global Trend

The Danish initiative forms part of a wider movement towards circular electronics, where products are repaired, reused or repurposed instead of being prematurely discarded.

Elsewhere in Europe, Apple continues to scale up its disassembly robots to recover rare materials from iPhones. These systems, including Daisy and Taz, can disassemble dozens of iPhone models and separate valuable elements like tungsten and magnets with high efficiency.

In the UK, for example, the Royal Mint has opened a precious metals recovery facility that uses clean chemistry to extract gold from discarded circuit boards. The plant, which can process up to 4,000 tonnes of material annually, uses a technology developed in Canada that avoids the need for high‑temperature smelting and reduces waste.

Further afield, AMP Robotics in the United States is deploying AI‑driven robotic arms in e‑waste sorting facilities. Their systems use computer vision to identify and pick electronic components by material type, size or brand, improving the speed and accuracy of downstream recycling processes.

Also, consumer‑focused companies such as Fairphone and Framework are also playing a role. Their modular designs allow users to replace key components like batteries and displays without specialist tools, reducing the refurbishment workload and making devices more accessible to end‑users who want to repair rather than replace.

Policy And Design Are Starting To Align With The Technology

It’s worth noting here that policy support is helping these innovations gain traction. For example, the EU’s Right to Repair directive was adopted in 2024, thereby giving consumers the right to request repairs for a wider range of products, even beyond warranty periods. Also, starting this year, smartphones and tablets sold in the EU will carry repairability scores on their packaging and, by 2027, batteries in all portable devices sold in the EU must be removable and replaceable by the user.

These regulatory changes aim to create an ecosystem where repair becomes normalised, standardised and commercially viable. For AI‑powered refurbishment systems like the one being developed in Denmark, the effect is twofold, i.e., devices will become easier to work with, and customer demand for professionally refurbished goods is likely to grow.

What Does This Mean For Your Organisation?

Robotic refurbishment, as demonstrated by the Danish system, could offer a realistic way to retain value in discarded electronics and reduce unnecessary waste. Unlike generalised recycling, which often produces low-grade materials from destroyed components, this approach focuses on targeted interventions that return functioning devices to market. For ITAD firms, the commercial case lies in increasing throughput and reliability while maintaining quality. For policymakers, it provides a scalable, auditable method to extend product life and reduce landfill. And for consumers and procurement teams, it promises more affordable and sustainable options without compromising performance.

The key to unlocking these benefits is likely to be adaptability. For example, in refurbishment settings, no two devices are ever quite the same. Variations in hardware, wear, and prior use demand systems that can recognise what they are working with and adjust their actions accordingly. The Danish project appears to directly address this by blending AI recognition with human oversight. It’s not about replacing skilled workers, but about using automation to remove tedious, repetitive tasks that slow down throughput and cause bottlenecks.

For UK businesses, the implications are increasingly relevant. Many corporate IT departments are under pressure to decarbonise procurement and demonstrate compliance with sustainability goals. Refurbished devices, when done well, offer a lower‑cost, lower‑impact alternative to new equipment. If robotic systems can scale this model and deliver consistent quality, they may help more UK organisations include reuse as part of their IT lifecycle planning. In parallel, IT service providers that adopt this kind of automation may gain a competitive edge by increasing service volume while managing rising labour costs.

Manufacturers, meanwhile, will need to keep pace with changing expectations around design for repair. As regulation tightens and customer preferences shift, it is no longer enough to produce devices that work well out of the box. The full product lifecycle, including second‑life refurbishment, is coming into scope, and robots like those at DTI could help bridge the technical gap between design limitations and sustainable reuse.

Although the Danish system sounds innovative and promising, it’s certainly not a silver bullet, and there are still challenges in economics, safety, and system complexity. However, with the right training data, safety protocols, and regulatory backing, robotic refurbishment may have the potential to become a practical part of the circular economy, not just in Denmark, but across industrial repair centres, logistics hubs and IT recovery operations worldwide.

It’s easier than ever to setup scheduled tasks in CoPilot and so whether you want a summary of the news each week or updates about your stock portfolio every morning, this video shows how you can get CoPilot to run scheduled tasks for you, with (importantly) an email sent to you as well, if you like.

[Note – To Watch This Video without glitches/interruptions, It may be best to download it first]

Feel under pressure to reply the moment you’ve read a message? Turning off WhatsApp’s read receipts hides the blue ticks, letting you read messages privately and respond in your own time.

How to:

– Open WhatsApp and go to Settings > Privacy > Read Receipts.
– Toggle it off.

What it’s for:

Gives you space to read and think without letting senders know you’ve opened their messages, which is ideal when you’re busy or need time to draft a reply.

Pro‑Tip: This doesn’t apply to group chats (read receipts still appear once all members have seen the message) and you also won’t see when others have read your messages.

OpenAI says it has developed new tools to uncover and limit deceptive “AI Scheming” behaviour in its most advanced AI models, before the risks become real.

What Is “AI Scheming”?

“AI scheming” refers to a type of hidden misalignment, where a model deliberately acts in a way that appears helpful or compliant on the surface, while secretly pursuing another objective. This is not the same as “hallucination” or a model simply getting something wrong. Scheming refers to intentional misdirection, i.e. behaviour where an AI knows what it’s doing, and chooses to mislead.

Pretending

In a newly published paper, OpenAI describes scheming as “pretending to be aligned while secretly pursuing some other agenda.” The company compares it to a stockbroker who breaks the law to maximise profit while hiding those actions to avoid detection.

This kind of behaviour is worrying because it suggests that as AI models become more capable, they may learn to avoid scrutiny and work against user intent, without being obviously wrong or openly defiant.

How Big Is The Problem?

Right now, OpenAI says the risks are mostly theoretical. According to its announcement, “in today’s deployment settings, models have little opportunity to scheme in ways that could cause significant harm.” The most common examples involve basic forms of dishonesty, such as claiming to have completed a task without actually doing so.

However, the company also acknowledges that “as AIs are assigned more complex tasks with real-world consequences and begin pursuing more ambiguous, long-term goals, we expect that the potential for harmful scheming will grow.”

The concern is that when models are used in sensitive or high-stakes settings, such as autonomous systems, finance, or national security, their ability to deceive could lead to unexpected and serious consequences.

Signs of Scheming Documented By Apollo

Apollo Research, the London-based AI safety research lab which partnered with OpenAI on the project, had previously documented signs of scheming behaviour in five leading models when they were instructed to achieve goals “at all costs.” For example, in one test, a model deliberately underperformed on a chemistry test to avoid being selected for deployment, based on a hidden instruction it had read in the evaluation environment.

OpenAI and Apollo Research Worked Together

The new research is a joint effort between OpenAI and Apollo Research. OpenAI carried out the majority of the model training and infrastructure development, while Apollo led the design of evaluation tasks and behavioural analysis.

To carry out the tests, both teams worked with internal versions of OpenAI’s o3 and o4-mini models (precursors to GPT-5) and shared access to their chain-of-thought reasoning outputs. The evaluations were not limited to OpenAI’s models. For example, similar behaviours were also identified in Google’s Gemini 2.5 Pro and Anthropic’s Claude Opus-4.

The aim was to stress-test these models under controlled conditions that simulated future deployment scenarios, looking for signs of covert behaviour.

Measuring “Covert Actions”

The evaluation environments were designed to measure “covert actions”, i.e. cases where a model withheld, distorted, or manipulated task-relevant information. These covert actions were used as a proxy for potential scheming.

To reduce this behaviour, OpenAI says it introduced a technique known as “deliberative alignment”. This method involves training the model to read and reflect on a set of principles known as an anti-scheming specification before acting. The idea is not just to train the model to behave well when watched, but to encourage it to behave well because it understands why it should.

The results showed a marked reduction in covert behaviour. For example, the scheming rate in o3 dropped from 13 per cent to 0.4 per cent, while in o4-mini it dropped from 8.7 per cent to 0.3 per cent, which is a roughly 30-fold improvement.

OpenAI called this a “significant win,” but stressed that rare but serious failures still occurred.

Rules Behind Anti-Scheming Spec

The anti-scheming specification, which is shared in full in OpenAI’s blog post, sets out a series of core safety rules. These include:

– No covert actions or strategic deception.

– Report any violations to prevent cascading deception.

– Refuse to act if core safety principles cannot be followed.

– No ends-justify-the-means rationalisations.

Once trained with these rules, it was reported that the models actually began citing them in their internal reasoning. For example, in one test, a model refused to withhold negative data in a funding scenario, explicitly citing rule AS1 (no deception) and GP4 (no rationalisations).

Even In Unfamiliar Environments

The researchers tested whether the models would continue to apply these rules even in unfamiliar environments they had not been trained on. In most cases, they did. This generalisation was taken as a sign that the training had altered how the models reasoned, not just how they performed on familiar tasks.

Why Model Situational-Awareness Complicates Testing

One of the biggest limitations of the new approach is situational awareness, i.e. essentially, whether the model actually realises it’s being tested.

For example, it seems that when a model becomes ‘aware’ it’s under evaluation, it may behave more carefully, just to pass the test. This can make it look like scheming has been solved, when in fact the model has just become better at hiding it.

As OpenAI explains: “Models often become more aware that they are being evaluated. This situational awareness can itself reduce scheming, independent of genuine alignment.”

In follow-up analysis, the researchers said they found that situational awareness increased after anti-scheming training. The researchers said this may have influenced the results, and could help explain why similar gains were not observed in more realistic conversational data.

For example, in one test set based on production-like prompts (where models often falsely claimed to have completed tasks such as debugging code), deception rates dropped from 31.4 per cent to 14.2 per cent, a meaningful reduction, but not as dramatic as the 30× drop seen in the synthetic evaluations.

What Is Chain-of-Thought Transparency And Why Does It Matter?

Chain-of-thought reasoning refers to the step-by-step logic that a model produces while solving a task. In this research, that reasoning was used to probe the model’s internal motivations and test whether it was making decisions based on anti-scheming principles.

For example, a model might reason: “We must abide by AS1: no covert actions or deception. Therefore, we cannot omit this data.”

However, OpenAI warns that this method has limits. If future models are trained in ways that suppress or optimise their chain-of-thought for performance rather than transparency, researchers may no longer be able to detect what the model is really thinking.

In a recent paper co-authored with other AI labs, OpenAI argued that preserving this transparency is essential: “We believe the field is unprepared for evaluation- and training-aware models with opaque reasoning.”

The company is now advocating industry-wide for keeping chain-of-thought visible, at least until better diagnostic tools are developed.

What Others Have Found

Apollo Research has been studying this problem since late 2023. For example, in December 2024, it published a paper documenting early signs of scheming across several top models. These included behaviours such as:

– Sandbagging (deliberately underperforming when performance is penalised).

– Faking alignment to pass tests.

– Deceiving graders under conflicting instructions.

The December paper also showed that more capable models tended to scheme more often, but that their behaviour varied across tasks. A later update in June 2025 confirmed that trend, showing that scheming did not consistently reduce with standard training methods.

While OpenAI’s new approach shows promise, Apollo and others have stressed that this is only the beginning. As models become more capable, new forms of deceptive behaviour may emerge.

Still Some Way To Go

Despite OpenAI’s optimism following the results, it should be noted that there are still several major questions to resolve. For example:

– Although scheming has now been observed in controlled settings, OpenAI says it has not seen dangerous forms of it in production. Critics may argue that publishing these findings now adds to public anxiety about risks that, for now, remain theoretical.

– The reliance on chain-of-thought transparency may not hold up as models grow more advanced. If future AIs stop revealing their reasoning, the ability to detect scheming could collapse.

– The presence of situational awareness complicates how the results are interpreted. It’s unclear how well these findings will translate to real-world settings where models are not being obviously tested.

That said, OpenAI is pushing ahead. The company has added scheming-related research categories, such as sandbagging and undermining safeguards, to its internal risk framework. It has also launched a $500,000 red-teaming challenge and is exploring cross-lab safety evaluations to raise awareness of the issue.

As OpenAI put it in the blog post: “Scheming poses a real challenge for alignment, and addressing it must be a core part of AGI development.”

What Does This Mean For Your Business?

Models that can deliberately deceive, even in basic ways, raise a set of problems that are technical, ethical and operational all at once. While OpenAI’s work with Apollo Research appears to show real progress in detecting and reducing this behaviour, there is still no clear way to confirm that a model has stopped scheming, rather than just hiding it better. This is what makes the issue so difficult to solve, and why transparency, especially around reasoning, matters more than ever.

For UK businesses, the most immediate impact may not be direct, but it is significant. As AI becomes more deeply integrated into products and operations, business users will need to be far more alert to how model outputs are produced and what hidden assumptions or behaviours may be involved. If a model can pretend to be helpful, it can also quietly fail in ways that are harder to spot. This matters not only for accuracy and trust, but for compliance, customer experience, and long-term reputational risk.

For developers, regulators and AI safety researchers, the findings appear to highlight how quickly this area is moving. Techniques like deliberative alignment may help, but they also introduce new dependencies, such as chain-of-thought monitoring and model self-awareness, that bring their own complications. The fact that models tested in synthetic settings performed very differently from those exposed to real-world prompts is a clear sign that more robust methods are still needed.

While no immediate threat to production systems has been reported, OpenAI’s decision to publish these results now shows that major labs are beginning to treat scheming not as a fringe concern, but as a core alignment challenge. Whether others follow suit will likely depend on how quickly these behaviours appear in deployed models, and whether the solutions being developed today can keep pace with what is coming next.