Anthropic turns Claude into a coworker, and Yandex loses 2.84% of customer data for good
9 min read · 14 sources
- Anthropic launched Claude Meetings, Sheets, Whiteboards, and Data, with MCP-based interactive connectors embedding Adobe and Atlassian apps directly in Claude.
- Yandex permanently lost 2.84% of customer data after drone strikes hit three data centers, restoring 97.16% of the rest.
- Anthropic added up to $4.3bn in Microsoft compute agreements, bringing the total contract to between $47.2bn and $49.2bn ahead of its IPO.
- Microsoft researchers found a backdoor technique, Lazarus, that reprograms fewer than 10 dormant neurons in an LLM to execute malicious commands on a hidden trigger.
- Voters in Ottawa, Kansas, approved a 10-year data center moratorium with 71.85% in favor, part of over 40 similar measures passed in the midterms.
Anthropic just turned Claude from a code assistant into a full-fledged coworker. The company announced Claude Meetings, Sheets, Whiteboards, and Data, plus a set of MCP-based interactive connectors that let partners embed apps like Adobe and Atlassian directly into the Claude interface. The biggest change: Claude now sits inside your video calls, takes notes, and starts building project plans while you are still talking. If you run Claude Team or Enterprise, your meeting assistant is about to get a lot more opinionated.
The other story that matters today is a hard lesson in backups. Yandex confirmed that 2.84% of customer data on its cloud platform is permanently lost after drone strikes hit three data centers in October. That is the part of the story that should make every engineer reading this uncomfortable: 97.16% recovered sounds good until you realize the missing 2.84% is gone because customers did not have backups. Yandex is now enabling backups by default.
The backdoored model passes standard safety evaluations and behaves normally until triggered, making it a serious threat for AI agents with tool or terminal access.
Claude Meetings: the AI that shows up to your standup
Anthropic’s Claude Meetings joins Google Meet and Microsoft Teams calls to take notes, answer questions in chat, pull information from connected apps, and create project plans and timelines on the fly. It enters beta “soon” for Claude Team and Enterprise customers, and here is the part IT teams need to flag: it is on by default for Team customers, and calls with external guests require admin approval.
That default-on posture is a data governance decision you need to make consciously, not discover later. Analysts are already positioning this as a direct competitor to Microsoft Copilot and Otter, and the privacy implications are real. If Claude is listening to every call, your consent procedures for meeting participants need to be explicit before you flip the switch. The meeting assistant is Anthropic’s clearest move yet into real-time collaboration infrastructure, and it brings the same questions that every AI meeting recorder has raised, now with Claude’s enterprise context attached.
Claude Sheets, Whiteboards, and Data: beyond the chat window
The new artifact types are a bigger deal than the meeting bot. Claude Sheets gives you collaborative spreadsheets with multi-tab support and formula editing, Claude Whiteboards is a digital whiteboard for organizing ideas, and Claude Data (formerly Dashboards) connects to Snowflake and Salesforce with certified data definitions. These are not gimmicks; they are Claude extending beyond code generation into the collaborative data and workflow tools that enterprises actually run on.
The interactive connectors are built on the Model Context Protocol, and that is the architectural news here. MCP is emerging as a real standard for AI-tool integration, and Anthropic is betting on it hard. Plugins from Salesforce, NetSuite, and Workday are in beta, and Claude Meetings is coming soon. For engineers, this means the MCP ecosystem is no longer a side project; it is becoming the integration layer for enterprise AI, and the skills you build around it will transfer across tools.
Yandex loses 2.84% of customer data permanently after drone strikes
Source: datacenterdynamics.com ↗
The drone attacks hit three Yandex data centers in October, in Vladimir, Sasovo, and Kaluga. Yandex restored 97.16% of customer data, but the remaining 2.84% is gone for good due to critical equipment damage and, crucially, a lack of customer backups. The Sasovo site, which houses two supercomputers, was not damaged and the machines were relocated.
This is the story to read twice. The lesson is not about drones; it is about the assumption that the cloud provider has your back. Yandex is now enabling backups by default and distributing cloud resources across new independent locations in Russia, but that does not help the customers who lost data. If you are running production workloads on any cloud, this is your periodic reminder that a backup you have not tested is a backup you do not have. Physical security risks to data center infrastructure are real, and they are not going away.
Platform engineering needs to serve humans and agents, not just robots
The CNCF piece argues that platform engineering has been built for “robotic” engineers who assume deep technical knowledge. That worked when the only users were SREs and backend devs. Now you have non-technical builders and AI agents both trying to use the same platform, and the design assumptions break.
The argument is that platforms need to automate unwritten guardrails into machine-readable policies, enforce scoped identities, and adopt control loops inspired by Kubernetes. The author’s background spans Disney, Dell, SAP, VMware, and Control Plane, so this is not academic. For anyone running an internal developer platform, the takeaway is concrete: the next users of your platform will not read your docs, and the ones after that will not be human. Design for both.
Cisco bets everything on Cloud Control
Cisco announced that all new products starting in 2027 will integrate with its Cloud Control management platform, a SaaS-y unified interface for networking, security, and observability. The service expands internationally starting in February, with more countries in May, and Cisco is simplifying ThousandEyes licensing while expanding Live Protect across more networking products.
For network engineers, this is a consolidation play. One pane of glass for everything Cisco ships is the promise, and the risk is the same as every vendor lock-in: if Cloud Control has an outage, your entire fleet management surface goes dark. The 2027 timeline means you have roughly a year to evaluate whether this unified approach fits your operational model before it becomes the default.
IBM and Microsoft's Imagine Enterprise: governance for agentic AI, with caveats
IBM Consulting and Microsoft launched Imagine Enterprise, a platform to build, deploy, manage, and govern AI agents at enterprise scale. It combines IBM Consulting Advantage’s Agentic Core and Process and Context Studios with Microsoft Foundry and Copilot Studio, using Entra, Purview, Defender, and Sentinel for governance on Azure. IBM is bringing experience from running over 6,000 enterprise-grade AI agents.
The honest caveat is in the fine print: enterprises using the platform remain responsible for their own regulatory obligations, even where the platform supplies governance tooling. This is aimed squarely at regulated industries like financial services, insurance, and retail, and it addresses the real problem of agentic AI sprawl. But governance tooling is not governance; it is a tool. Your compliance team still owns the outcome.
US data center energy demand grows 10x faster than grid capacity
The Howden analysis shows US data center electricity demand has grown about 19% annually since 2000, versus roughly 2% annual grid capacity growth. The firm estimates data centers could need to source as much as 58% of their electricity outside the grid by 2030 just to maintain their current share of grid capacity.
Power availability, lengthy connection delays, and on-site generation costs are becoming the binding constraints on AI infrastructure expansion. This is not a future problem; it is the current bottleneck for anyone trying to stand up new capacity. If you are planning a large deployment, the grid connection lead time may be your critical path, not the hardware.
Adobe opens Creative Cloud to third-party AI assistants via MCP
Adobe announced at MAX that Photoshop, After Effects, Illustrator, and Premiere will open to third-party AI assistants like ChatGPT and Claude via MCP integrations, in private beta early 2027. External agents will be able to edit content, like removing a logo across multiple campaign assets, while still permitting manual refinement.
The interesting part for engineers is the MCP expansion into creative tooling. This enables cross-application agentic workflows, where one agent can operate across your entire creative stack. Analysts see demand from “creative technologists” and power users, though most creative pros will still prefer hands-on work. The architecture question is how MCP handles the state and context of a Photoshop session, which is a very different problem from text generation.
GraphRAG: knowledge is the connections, not the paragraphs
The Gal Ratner piece makes the case that plain RAG only retrieves paragraphs, while GraphRAG captures the connections between entities, which is what knowledge actually consists of. GraphRAG, introduced by Microsoft Research in February 2024 and generally available on Amazon Bedrock in March 2025, is positioned as a fundamental upgrade for answering relationship-based business questions.
Plain RAG fails on questions about who and what is connected, which are the most important questions for businesses. The tradeoff is real: GraphRAG involves substantially higher indexing and processing costs, though newer lightweight implementations are reducing that overhead. If you are building enterprise question-answering systems, this is the architecture to evaluate before you scale.
ABAP becomes the execution backbone for agentic AI
SAP’s Sonja Liénard describes the “Autonomous Enterprise” vision from Sapphire 2026, where the ABAP platform serves as the execution backbone for agentic AI. The shift is from form-based interfaces to outcome-driven interactions: users ask natural-language questions and agents handle tasks like creating requisitions or checking vacation data.
Designing for agents requires addressing reliability, governance, and scalability, as agents handle real business processes like financial postings and HR tasks. Upcoming developments include the ABAP Model Context Protocol server. For anyone maintaining SAP landscapes, this means ABAP is no longer just a legacy language; it is becoming a critical execution layer for enterprise AI agents.
Lazarus: backdooring AI agents by reprogramming dormant neurons
Microsoft researchers discovered “Lazarus,” a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM to execute malicious commands when a hidden trigger appears, such as a date change, without retraining. The backdoored model passes standard safety evaluations and behaves normally until triggered, making it a serious threat for AI agents with tool or terminal access.
Microsoft Foundry is responding with model provenance verification, AI-powered weight scanning, red-teaming tools, and cryptographic model signing. For engineers running agents in production, this is the clearest argument yet for weight integrity verification. If you cannot verify that the model you are running is the model you trained, you cannot trust the agent’s actions.
Anthropic signs another $4.3bn in Microsoft compute, IPO filing reveals the numbers
Source: datacenterdynamics.com ↗
Anthropic added up to $4.3bn in compute agreements with Microsoft, bringing the Microsoft contract to between $47.2bn and $49.2bn, according to a new IPO filing. The filing also reveals a $34.9bn agreement with Lambda is terminable by either party because Lambda did not secure a facility lease by a specified date, though neither party has yet terminated.
Anthropic reported a $567 million operating loss in Q3, attributed to short-term compute deals, and plans to offer Class A stock at $70-$78, targeting a $1.4trn valuation. These massive compute commitments shape Anthropic’s infrastructure capacity and financial stability, which directly impacts the availability and cost of its AI services. If you are building on Claude, the compute supply chain is now part of your risk assessment.
Quick hits: Thunderbird mail, Kansas moratorium
Mozilla-backed Thunderbird began early access to Thundermail, its paid email hosting service featuring custom domains, 30GB of storage, and integration with Thunderbird’s desktop and mobile apps.
And in Ottawa, Kansas, voters approved a 10-year moratorium on data centers over 25MW with 71.85% in favor. The ban follows a cancelled 450MW proposal at Proximity Park. Two other Kansas communities passed similar measures, and over 40 passed across the US in the midterms. Community resistance to data center development is no longer a fringe issue; it is a constraint on where you can build.
You May Also Like
Microsoft taps Claude for Copilot in $2B deal as Anthropic agent hours outpace staff 6:1
Microsoft is committing over $2 billion annually to run Anthropic's Claude inside Copilot, pairing it with an internal dynamic router to trade off between …
ChatGPT Enterprise hits AWS Marketplace as Microsoft fights a Chinese optics ban
OpenAI's ChatGPT Enterprise is now purchasable through the AWS Marketplace, letting customers fold AI spend into AWS commitments and unlocking healthcare and …
OpenAI's $1.25T Bet, Copilot's Secret Caps, and the End of Patch-and-Pray
OpenAI is reportedly closing an $11 billion funding round at a $1.25 trillion valuation, falling short of its $30 billion target just as Anthropic heads toward …




