7 min read · 14 sources

Your employees are paying for AI tools out of their own pockets, and they are not telling you. A new Deloitte UK report says 63% of British workers now use generative AI at work, but 31% are doing it without their employer’s knowledge and 17% are personally footing the bill. That is £958 million of ungoverned software already running inside your network, and it is not going to stop because you ask nicely.

The real story is not the spending - it is the exposure. Nearly half of those workers have had no formal safety training, and 64% worry their managers think AI can do their jobs. That combination means your data is being piped into third-party models with no oversight, no logging, and no kill switch. Shadow AI is not a trend; it is the default state of your enterprise, and it is already inside systems you thought you controlled.

63% of British workers use generative AI at work, and 17% pay for it themselves - £958 million in ungoverned shadow IT.

Salesforce and Google Drive Both Went Down on the Same Day

Source: itpro.com ↗

Salesforce took a widespread service disruption on September 16, knocking out core services for hours while Dreamforce 2026 was ramping up. Google Drive separately suffered a 37-minute disruption the same day, a reminder of how much employee productivity now depends on a small number of SaaS platforms and why IT still needs outage communications and fallback plans.

The Salesforce incident is a case study in single-vendor risk: when your CRM, marketing automation, and support desk all live in one cloud, a bad config or a failed deployment takes down your revenue operations with it. Engineers should be asking hard questions about blast radius and whether their runbooks actually cover third-party dependencies.

Gemini in Google Workspace Can Now Connect Directly to Salesforce, HubSpot, and Asana

Source: workspaceupdates.googleblog.com ↗

Google is rolling out MCP-based integrations that let Gemini interact with Salesforce, HubSpot, Asana, and other Workspace apps. The announcement covers connectors that reduce friction in automating tasks across disparate SaaS systems, potentially streamlining data flow and process automation.

IT can control these integrations by domain, organizational unit, or group, so this is not just another end-user AI feature - it is something Workspace admins need to review before enabling. Specific technical details like version numbers or pricing tiers (e.g., Google AI Pro, Ultra) are mentioned but not elaborated, so treat the rollout as a pilot, not a finished product.

Not Every App in an Enterprise Is an Enterprise App

Source: blog.mastykarz.nl ↗

Waldek Mastykarz makes a sharp argument in his blog post: the defining factor for whether software is an “enterprise app” is business dependency, not its technology, builder, or procurement model. A simple script that generates invoices is enterprise software if the business relies on it to operate; a $10 million platform is not if nobody depends on it.

Engineers should use this dependency-based definition to prioritize monitoring, support, and continuity planning for systems that truly matter. Stop over-engineering for low-impact tools and start treating business-critical shadow apps with the same rigor as your sanctioned stack - because if it breaks and the business cannot operate, nobody cares who bought it.

Gartner: AI Will Remake Business in the Next Three Years

Source: itpro.com ↗

Gartner predicts AI will introduce new business challenges within three years, including physical AI (robots), agent swarms clogging public services, and difficulty managing AI costs and risks. The report specifically calls out the need for real-time AI cost controls and guardrails that actually exist and work.

Engineers should prepare for AI-specific attack surfaces and resource management, integrating monitoring into their security and cost strategies. If you think your Prometheus metrics cover GPU spend, think again - agent swarms will burn through tokens and compute in ways that look nothing like traditional workloads.

Chrome 155 Kills Debugging Extensions Under Enterprise Policies

Source: developer.chrome.com ↗

Starting with Chrome 155, extensions that use the chrome.debugger API will be blocked from attaching when enterprise policies like runtime_blocked_hosts or DisableScreenshots/DLP are set. According to the Chrome developer blog, chrome.debugger.attach() will be rejected with specific error messages, even if allowed origins are listed - an all-or-nothing model that addresses CDP’s inability to use origin-based filtering securely.

The change only affects managed browsers with those policies set, but some extensions may stop working unless they move to more limited APIs. If you rely on debugging extensions in enterprise environments, update your error-handling code to check for these new policy-related messages and plan for stricter environments.

Rivet BYOC Lets You Run the Control Plane Inside Your Own VPC

Source: rivet.dev ↗

Rivet released BYOC (Bring Your Own Cloud), deploying its control plane inside a customer’s AWS or Google Cloud VPC, managed by Rivet. Data and compute stay within the security perimeter, with only outbound connections to Rivet Cloud, and each region gets its own VPC, cluster, and control plane. The changelog notes you can expose Rivet over HTTPS on your own hostname or keep it private.

This is a big deal for regulated industries: your data never leaves your account, but Rivet handles deployments, updates, and maintenance via an operator running in a Kubernetes cluster. If you have been holding back on edge deployment because of data residency, this removes that excuse.

Teach Gemini Your Team's Know-Hows with Skills in Google Workspace

Source: workspace.google.com ↗

Google is rolling out Skills in Workspace, letting organizations feed Gemini their prompts, rules, templates, and reference material so it can repeatedly follow the same process across Gmail, Docs, and other apps. The announcement cites examples like standardizing templates and automating invoice reviews.

Teams can collaborate on skills like documents, while centralized governance and Admin console controls keep things in check. This is effectively a low-code way to codify tribal knowledge, but it also means your prompt engineering becomes a maintained asset - version it, review it, and treat it like code.

Salesforce Evolves Well-Architected Framework for the Agentic Enterprise

Source: salesforce.com ↗

Salesforce updated its Well-Architected Framework to address the Agentic Enterprise, where architects design systems with autonomous agents and interconnected experiences. The framework provides principles for balancing business needs, technical requirements, and architectural quality, with dedicated sections on Performance, Cost, Optimization, and Fairness, each featuring a dedicated Agentic section.

It will be unveiled at Dreamforce 2026. Engineers should adopt this to guide decisions on complex, less deterministic systems, ensuring trusted and sound architecture - because agents fail differently than deterministic services, and your SLIs need to account for that.

MCP Servers on Employee Devices Are an Endpoint Inventory Problem

Source: repello.ai ↗

The Repello article makes a critical point: MCP (Model Context Protocol) servers lack an admin console, with configurations stored in per-user JSON files on individual laptops, editable without approval or logging. This makes endpoint inventory the core issue, as server names alone do not validate or register anything.

To get a complete inventory, security teams must read per-client configuration files, including paths, versions, and environment variables in use. This is not an identity problem or a network problem - it is a device management problem, and your EDR probably does not cover it yet.

Dreamforce 2026: AI Safety Is the New Big Tech Battleground

Source: itpro.com ↗

Dreamforce 2026 focused on AI safety as a battleground, with leaders like Amodei and Altman discussing the need to slow AI development and set industry standards. The coverage highlights a growing disagreement among AI leaders on how to proceed, with some pushing for brakes and others for full speed.

Regardless of which approach wins out, enterprises deploying agents will increasingly need their own standards for what models and capabilities they are willing to allow. One emerging idea is a SOC 2-like independent assurance layer for AI agents - watch this space.

AIUC Raises $40 Million to Certify Enterprise AI Agents

Source: securityweek.com ↗

AIUC raised $40 million to certify enterprise AI agents, according to SecurityWeek. The certification standards may indicate reliability and compliance, though the article does not detail the criteria.

Engineers should consider certification standards when selecting AI agents, but treat them as a baseline, not a guarantee. A $40 million round means the market is taking this seriously, but certification is not a substitute for your own testing and monitoring.

Best Knowledge Engine Platforms in 2026

Source: pinecone.io ↗

Pinecone published a review of knowledge engine platforms, distinguishing between full knowledge engines and lower-level tools like vector databases (Pinecone Nexus), graph databases (Neo4j), and RAG frameworks (LlamaIndex). It recommends Pinecone Nexus for curated, task-specific knowledge, but engineers should match platform choice to their specific data relationships and integration needs.

Do not buy a vector database because it is popular; buy it because your data model needs it. The differentiation between full engines and their underlying components matters more than the marketing.

Microsoft Patch Tuesday Breaks Copy-Paste in Excel

Source: itpro.com ↗

Microsoft’s September Patch Tuesday inadvertently took down copy and paste in Excel, and there is no way to fix it yet. The ITPro report notes Microsoft acknowledged the issue and is working on a fix, but no workaround exists yet.

If your finance team lives in Excel, this is an emergency. Pin the patch if you can, and warn users that Ctrl+C/Ctrl+V may silently fail - because the last thing you want is a spreadsheet full of stale numbers going into a board report.