OpenAI’s GPT-5.6 now in Microsoft 365 Copilot

photo by Microsoft

Microsoft has officially announced the availability of OpenAI’s GPT-5.6 family of models within Microsoft 365 Copilot – specifically in Word, Excel, PowerPoint, Copilot Chat, and Copilot Cowork. OpenAI and Microsoft partnered together to optimize GPT-5.6 for complex knowledge work across Microsoft 365 applications, making it the preferred model for Microsoft 365 Copilot.

Inside the GPT-5.6 lineup: from flagship to lightweight

The new model family has been designed to address diverse market needs:

• GPT-5.6 Sol – the flagship model built for the most complex tasks, such as advanced reasoning, coding, scientific work, cybersecurity, and long-running agentic tasks.
• GPT-5.6 Terra – the balanced intermediate model. According to OpenAI, it delivers performance competitive with GPT-5.5 while being twice as cheap.
• GPT-5.6 Luna – the smallest and most cost-efficient variant, aimed at faster, lower-cost workloads.

Along with Sol’s debut, OpenAI introduced a new “max reasoning effort” setting, which allows users to extract even better results from the flagship. A new “ultra mode” has also appeared, going beyond the capabilities of a single agent by utilizing subagents to accelerate complex workflows.

Higher-quality work across everyday apps

Daily tasks within Office apps are becoming more streamlined thanks to the stronger reasoning capabilities of the new architecture. GPT-5.6 helps Copilot turn rough intent into high-quality outcomes:

• In Word: it helps turn rough ideas into more complete drafts, organize content with stronger structure and flow, and refine writing into a polished artifact.
• In Excel: it helps reason through more complex analysis and move from task to outcome with less manual assembly.
• In PowerPoint: it enables the creation of richer presentation drafts with stronger slide content, better visual balance, and more flexibility across presentation styles.

In Copilot Chat, GPT-5.6 helps Copilot reason through more complex asks-comparing options, structuring plans, troubleshooting problems, and working through ambiguity-to turn open-ended prompts into clearer, more actionable responses.

Furthermore, GPT-5.6 strengthens agentic work in Copilot Cowork, carrying complex multi-step tasks from instruction to completed results. The user simply describes the desired outcome, and Cowork plans, reasons across tools and files, and runs the work end-to-end, returning a finished deliverable rather than just a draft or recommendation. Paired with Work IQ, Cowork brings the right context to each step so the output accurately reflects the company’s files, tools, and business environment.

Deployment and availability details

The rollout of GPT-5.6 to Copilot in Word, Excel, PowerPoint, Chat, and Copilot Cowork has already begun. Once available, Copilot may automatically use GPT-5.6 when it is best suited for the task. Where model selection is enabled, users can also choose GPT-5.6 directly from the model selector. Availability may vary depending on the region and tenant configuration.

The new models have also begun rolling out across OpenAI’s own products, including ChatGPT, Codex, and the OpenAI API. Users on the ChatGPT Free and Go tiers will gain access to the Terra version, while subscribers of Plus, Pro, Business, and Enterprise plans will have full access to choose between Sol, Terra, and Luna, along with the ability to adjust the reasoning effort level. Additionally, the new “max mode” is being deployed for ChatGPT Work and Codex users. The advanced “ultra mode” will be available in ChatGPT Work for Pro and Enterprise accounts, as well as in Codex for Plus and higher tiers.

Claude Sonnet 5 brings multi-step agentic AI to Copilot users

photo by Anthropic

Anthropic’s Claude Sonnet 5 has launched in Microsoft 365 Copilot, rolling out to Copilot Cowork and PowerPoint. This highly agentic model enables autonomous, multi-step business workflows across documents, spreadsheets, and presentations, backed by Work IQ. Delivering performance close to the flagship Opus 4.8 at a lower price point, Sonnet 5 is also available across all Claude plans and the API, featuring enhanced security safeguards and optimized cost efficiency for everyday enterprise tasks.

Availability in Microsoft 365 Copilot

Anthropic’s latest frontier model, Claude Sonnet 5, has officially joined the Microsoft 365 Copilot lineup, rolling out specifically to Copilot Cowork and Copilot in PowerPoint. The model is tailored for agentic, multi-step tasks across daily business activities, including document creation, spreadsheet management, and presentation design.

When integrated with Work IQ, Copilot utilizes Sonnet 5 to plan and build business artifacts grounded directly in user files, chats, meetings, and workplace data. Corporate IT administrators can manage access to the model based on internal organizational policies. Deployment schedules remain subject to regional and tenant configurations.

Capabilities and performance

Claude Sonnet 5 is designed as a highly capable agentic model that can autonomously formulate plans, utilize tools such as browsers and terminals, and execute workflows at a level previously reserved for larger systems. It fills the gap between standard models and top-tier performance, delivering results close to Opus 4.8 but at a lower price.

photo by Anthropic

The model marks a substantial improvement over its predecessor, Sonnet 4.6, in core areas like reasoning, coding, and tool use. Users can adjust the model’s effort levels to balance financial costs against performance requirements. Early testers reported that Sonnet 5 successfully completes complex pipelines where older iterations would stop short, while also independently reviewing its own output.

Safety evaluations indicate that Sonnet 5 is more secure than Sonnet 4.6, demonstrating greater resistance to prompt injections, malicious requests, hallucinations, and sycophancy. The model is not trained for advanced cybersecurity tasks and shows poor exploitation capabilities compared to Opus 4.8. To ensure safe operation, real-time cyber safeguards have been activated by default.

No more tool switching. Copilot now edits directly inside Word for iPad

photo by Microsoft

Microsoft has announced the expansion of a feature already familiar to desktop, web, and mobile users. Microsoft 365 Copilot is now available directly within the Word app on iPad, allowing seamless text collaboration without ever leaving the document.

New editing capabilities on tablet

iPad users can now treat Microsoft 365 Copilot as a personal partner in the creative process. The tool enables drafting new content, refining existing text, and reorganizing document structures directly on the page. Importantly, all modifications are made while preserving the original formatting and built-in Word styles.

This integration eliminates the need to constantly switch between different applications or commands. Copilot helps efficiently transform rough ideas into polished materials, handles much of the manual labor involved in editing, and allows authors to stay fully focused on their writing flow.

How to launch Copilot in Word for iPad

1. Open the Word app on your iPad and sign in with a Copilot-enabled account.
2. Create a new document or open an existing file, then select the edit option.
3. Tap the Copilot icon to open the Copilot Chat.
4. Enter your request in the prompt box.
5. Review and apply the changes suggested by the AI.

Practical tips and limitations

To get the best results, it is crucial to use specific prompts—instead of a general request to update a file, it is better to point to a specific data source, such as an email from a particular month. To reference external files, meetings, or messages, users can simply type “/” in the chat box or use the source picker. If editing is not enabled, it can be activated using the plus button in the chat menu.

It is worth noting that Copilot works directly on the open file, but it displays a preview of proposed changes for approval when working in shared documents. Changes can be undone at any time using the Version History in OneDrive or SharePoint. Additionally, Copilot respects the Track Changes feature, but it cannot create entirely new files from scratch, add or edit comments (and modifying text may remove associated notes), or generate images while in editing mode—to create visuals, users must turn off editing first.

Scenarios to try

The new feature can be tested in various ways, such as asking the AI to draft a one-page project update based on existing notes, including status updates and risks. The tool also excels at rewriting text to sound more confident and customer-ready, or reorganizing a document to improve scannability by adding clear headings and an executive summary.

This new capability is rolling out to users with access to Microsoft 365 Copilot. To utilize this feature on an Apple tablet, users must be running Word for iPad Version 2.109 (Build 26051516) or later.

Microsoft uses AI to shield Windows against zero-day vulnerabilities

photo by Microsoft

With AI accelerating the ability of attackers to find and exploit security flaws, Microsoft is integrating advanced AI models into the Windows development and review pipeline. This shift aims to identify code patterns, prioritize risks, and uncover vulnerabilities before they can be weaponized as zero-day exploits.

Inside MDASH: scaling vulnerability discovery with multi-model AI

The core of this strategy is MDASH (Microsoft Security’s multi-model agentic scanning harness), which operates on dedicated cloud infrastructure:

• The Scanner Pipeline: uses a multi-model “debate” among various AI families to scan critical binaries and flag potential vulnerabilities.
• The Prove Pipeline: a Windows-specific secondary stage that validates these candidates to eliminate false positives, ensuring only high-confidence findings reach human engineers.

This automation allows Microsoft to update its Secure Development Lifecycle (SDL) to actively counter AI-enabled attack vectors during the initial building phase.

Accelerating remediation and fix validation

AI is also being used to streamline the engineering process after a vulnerability is found. It compresses the timeline from discovery to a validated fix by helping developers:

• Understand system failures and code regressions much faster.
• Propose secure candidate fixes that match the structure of the surrounding codebase.
• Surface related vulnerabilities across other parts of the Windows estate.
• Automatically select the most relevant regression tests for the code change.

Microsoft is continuously investing in these Windows-specific tools and agentic harnesses to enable the end-to-end generation and validation of fixes. While AI automates the bulk of the discovery and testing, human expertise remains central to reviewing code, making risk-based decisions, and maintaining quality control.

Poland becomes third European nation to access OpenAI’s restricted GPT-5.5 Cyber model

photo by OpenAI

Poland has officially become the third country in Europe and one of a select few globally to secure access to OpenAI’s pioneering artificial intelligence model, GPT-5.5 Cyber. The breakthrough deployment was announced by the NASK National Research Institute (NASK-PIB). Developed specifically to detect, map, and deeply analyze cyber threats, the framework marks a monumental shift in the country’s national digital defense capabilities.

The solution will not be deployed on the commercial market. Access is strictly controlled and granted exclusively to trusted institutions via the specialized Government Trusted Access for Cyber program, designed solely for state agencies and critical security entities.

Advanced vulnerability remediation and benchmarks

GPT-5.5 Cyber is a highly specialized model engineered for advanced automated vulnerability hunting within complex telecommunication and information systems. Industry benchmarks indicate that the system’s analytical capabilities perform on par with Anthropic’s highly classified Claude Mythos 5 – a competing model that has recently been placed under a total foreign embargo by the United States government.

Crucially, OpenAI’s new tool does not simply locate vulnerabilities; it possesses the capability to autonomously generate secure patches and automatically remediate detected flaws in real time. It is also fully optimized for reverse-engineering and conducting exhaustive deep-dives into sophisticated malware code.

Gaining access to this advanced infrastructure opens unprecedented strategic capabilities for the CERT Polska team, which functions natively within the structures of NASK-PIB to mitigate active digital threats.

Augmenting rather than replacing human expertise

Identifying zero-day vulnerabilities and evaluating critical threat landscapes remains an incredibly time-consuming operation requiring highly scarce, specialized knowledge. While GPT-5.5 Cyber can instantly parse massive datasets, map potential attack paths, and surface anomalies hidden from human oversight, leadership stresses that the technology is meant to amplify human analysts rather than replace them.

Of course, this doesn’t mean we will suddenly stop needing specialists. I will always repeat that tools are just tools, and they will only function correctly in human hands. A human always stands at the end of the analysis conducted by the AI model, and a human verifies exactly what the tool has produced. While it is incredibly useful, it doesn’t mean we were any less effective without GPT Cyber – said Radosław Nielek, Director of NASK

The strategic deployment of this highly restricted tool also serves as a strong international nod toward the expertise and maturity of the Polish cyber-defense community. As the primary national response hub, CERT Polska will leverage this asset to handle critical incident telemetry and systematically shield the country’s digital borders against state-sponsored and high-tier cyberattacks.