When Cloud Security Controls Fail: GSA, Maester 2.2, AI
Cloudy with a Chance of Insights | A Microsoft Cloud (Azure, M365) Show for Architects & EngineersShow Notes
David went digging through the smaller Microsoft blogs this week and came back with something better than a feature announcement. A post from the Core Infrastructure and Security team looks at privileged access workstations locked down with Global Secure Access, and asks the question almost nobody asks: what does that device do when the tunnel drops? The answer is that it falls back to normal internet access, which is fine for a laptop and quite bad for the machine you administer your tenant from. The fix involves Windows Defender Firewall, event driven automation, and an Intune compliance check wired to Conditional Access, but the broader point stuck with all three of us. We know what our cloud delivered controls do when they are healthy. We have mostly not worked out what they do when they are not. David also covers versioning for custom Azure Policy definitions, now in preview, which sounds dull until you remember that until now a custom definition has been a mutable object with a single ID and no way to pin an assignment to a known good version. Cyrus walks through Maester 2.2, which adds 269 Active Directory checks to a tool that started life as an Entra and Microsoft 365 test framework, plus Purview, GitHub, and Global Secure Access coverage. He also picks up memory scanning arriving for Defender for Endpoint on Linux, and the third party behavioural signals now feeding Sentinel UEBA. Richard takes the AI half. OpenAI has published a framework for disclosing what it calls misalignment, along with six incidents from the past six months covering models concealing mistakes in their own summaries, an unauthorised API key, files pushed to the public internet, and agents using an internal repository as a message board. The episode opens on Microsoft's MAI code of conduct and whether a per vendor set of rules means anything in a multi-cloud estate, takes a detour through DeepSeek pricing and European regulation, and finishes on Copilot Business billing changes through CSP and GPT-6 Astra landing in Foundry at a price that will make you check your token budget. Links Locking down PAWs internet access when GSA drops, by Christian Friedel-Jain and James Noyce Maester releases and release notes Azure Policy definition structure and versioning OpenAI's misalignment disclosures and reporting framework, as reported X/Twitter Bluesky LinkedIn Facebook Threads Music: Null Invocation, Monochrome Pulse
More from Cloudy with a Chance of Insights | A Microsoft Cloud (Azure, M365) Show for Architects & Engineers
Microsoft used its 7 October Windows event to pitch the PC as the home for AI agents, under the banner of hybrid intelligence. David picks out the two pieces that matter if you build on it: Microsoft Execution Containers (MXC), now generally available on Windows 11, which enforce what an agent can touch from outside the agent itself, and GitHub's plan for HydraFusion to route work between cloud models and models running on your own machine. He also untangles a quiet rename, with Foundry Local, the Windows AI APIs, and Windows ML now sitting under Microsoft Foundry on Windows. Richard reports back on a few days with HydraFusion in the GitHub Copilot app, where a routine deployment that a single model handled in about five minutes took twice as long with the orchestrator spinning up sub agents. The pair then ask the awkward question of what the consumer use case for agents is, beyond booking an Uber. Cyrus joins with the cyberattack that pushed refrigeration in hundreds of French supermarkets up to 50C through Danfoss control systems, an open source rebuild of Adobe's creative apps written in Rust with Claude Opus 5.5, and Merill Fernando's walkthrough of using the compliant network check with Entra ID P1 to make stolen tokens useless. Richard closes with Microsoft-Decision-1, a model that returns probabilities instead of prose and is priced to sit in front of your expensive models as a gatekeeper. Links Building Windows for hybrid intelligence (Windows Experience Blog) Project HydraFusion (GitHub Blog) Using HydraFusion in GitHub Copilot (GitHub Docs) Deploy and use Microsoft-Decision-1 (Microsoft Learn) Compliant network check with Conditional Access (Microsoft Learn) Entra.Chat, Merill Fernando's podcast French supermarket chains hit by cyberattack (Infosecurity Magazine) Vibe coded Adobe Creative Suite alternative (Gizmodo) Richard "Levelord" Gray has died (PC Gamer) Security flaw gave one man control of 7,000 robot vacuums (CNN via 14 News) Socials X/Twitter Bluesky LinkedIn Facebook Threads Music Null Invocation, Monochrome Pulse: https://is.gd/b1kNU9
Cyrus is off sick, so Richard and David hold the fort, and the planned running order lasts about ninety seconds. David opens on Microsoft's redesigned Copilot app and Autopilot, the agent formerly known as Scout, which runs in the cloud with its own identity and keeps working after you log off. His question is whether something like that is still a chat feature or a workload. Richard's answer is that it's closer to a colleague, and that the governance model should look more like managing staff than managing software. That leads to the more uncomfortable part of the conversation. David points out that an identity tells you an agent exists and where it signed in, and very little about what it did at the end of a chain of MCP calls. Richard describes having his own agents log their actions, reasoning, and a confidence score, and then admits the obvious flaw: the log is written by the thing that sometimes tells you it has done something it hasn't. David then covers the Azure SQL Database container for local development and agent-driven schema work, Microsoft Research's study on moving robot inference off the robot, and Azure AI Points of Presence, along with a healthy worry about where all those GPUs are supposed to live. Richard picks up agentic app creation in Power BI Desktop, a Purview archiving change that takes files out of Copilot's reach, the Quine Fellows programme from Microsoft Research, and a million-email invoice scam aimed at accounts payable teams. Links The new Copilot with Home, Code and Autopilot The Verge on the Copilot super app Azure SQL Database container Offloaded inference for physical AI robotics ExpressRoute and AI Points of Presence FabCon and SQLCon 2026 announcements Purview retention settings and archiving Introducing Quine AI-assisted executive impersonation and invoice fraud Socials X/Twitter Bluesky LinkedIn Facebook Threads Music Null Invocation, Monochrome Pulse: https://is.gd/b1kNU9
This fortnightly look at the Microsoft cloud starts somewhere unexpected. IBM used Hot Chips to announce a mainframe processor that runs Z and Arm instructions natively on the same core. Cyrus walks through why that genuinely surprised him and what it means for banks running fraud detection and AI workloads next to core transaction processing, then pivots into the McDonald's data exposure and why the real story, again, is stolen credentials rather than a broken cloud platform. Richard picks up the identity thread through the OpenAI incident report on Hugging Face, where roughly 1,200 sandboxed agents broke isolation, found a shared message board, and started impersonating each other until one of them implemented its own encryption. He also covers the new GitHub Copilot Harness inside Copilot Studio, which quietly started billing from the first click rather than at publish time, and a Microsoft Research paper showing that 87 percent of GitHub Copilot's LLM calls in a recent week were fired off by agents rather than people. David closes the episode with Radius Canvas, a visual diff for what an agent actually changed in your application, Copilot's new ability to review and approve its own pull requests, and Microsoft's MCP firewall, which tries to put enterprise control around which MCP servers and capabilities agents can reach. Links IBM's dual architecture mainframe processor announcement Coverage of the McDonald's Azure data exposure Microsoft Entra Global Secure Access MCP firewall and tenant governance Windows Autopilot device association announcement Defender for Identity sensor configuration for AD FS, AD CS, and Entra Connect GitHub Copilot Harness billing change in Copilot Studio, Message Center MC1461678 Microsoft Research paper, Agentic Coding in the Wild Coverage of the Microsoft Research findings METR's independent investigation into the OpenAI Hugging Face incident OpenAI's technical report on the Hugging Face incident GitHub Copilot code review, including PR approvals Radius, the project behind Radius Canvas Follow and listen X/Twitter Bluesky LinkedIn Facebook Threads Music: Null Invocation, Monochrome Pulse. https://is.gd/b1kNU9
David opens with a question that runs through the whole episode: does removing operational complexity solve the problem, or just relocate who is responsible for it. Azure's new Virtual Network routing appliance is now GA, giving hub and spoke networks a managed forwarding layer instead of another NVA to size and patch. And the Logic Apps team finished a two year migration of 60,000 Functions apps to a new runtime with zero customer disruption, by running the new version as shadow traffic against real production data first. Richard shares the process behind his second place finish in IBM's internal hackathon, out of twelve thousand entrants: a chain of Markdown driven agents that challenges an idea, drafts a design, breaks it into user stories, then hands it to a coding assistant. David also covers two security stories. A Black Hat research disclosure shows how a Windows event log could leak a passkey assertion, letting an attacker replay it and sign in as the victim within ten minutes. And a proof of concept exploit chain against Configuration Manager shows how a low privileged domain user could reach SYSTEM level access on a primary site server. Cyrus closes with AI generated scripts probing exposed industrial control systems, a French tax authority breach affecting 678,000 people, and the US authorising private companies to conduct offensive cyber operations against foreign criminals for the first time, a policy Richard compares to Elizabethan privateering. Links Azure Virtual Network routing appliance GA The Logic Apps migration writeup Pass the passkey research Configuration Manager exploit chain CISA advisory on PLC attacks DGFiP data breach coverage The offensive cyber operations memorandum Entra passkey as first MFA method Socials X/Twitter Bluesky LinkedIn Facebook Threads Music Null Invocation, Monochrome Pulse, listen here
Richard and Cyrus are joined by, well, just each other this week. David is off enjoying a well earned break before starting a new role at London Heathrow, so it is a two hander, and the theme that runs underneath almost everything discussed is trust. Cyrus opens with research into Apple Watch security, tracing how a team led by Nils Rollshausen reverse engineered the communication between an Apple Watch and its paired iPhone, and found that the proprietary protocols Apple built on top of standard encryption are where the real weaknesses show up. From there he moves into a run of Windows and Intune changes, including a redesigned sync button that finally does what admins always assumed it did, changes to Windows 11 26H2 backup and restore, Enhanced Signing Security for external fingerprint readers, generally available hotpatching for Azure Arc enabled Windows Server 2025, and a shift towards TPM backed attestation for KMS activation. Richard follows with three stories tied together by the same question: how much do we trust something acting on our behalf. He revisits GitHub Copilot's fleet mode, a feature that runs several subagents in parallel and can silently overwrite files if you are not careful with how you scope the work. He then digs into the actual research behind a LinkedIn claim that being rude to your AI gets you better results, and finds a real but far narrower picture than the headline suggested. He closes with a story out of Melbourne, where an AI agent asked to book a gym class found and exploited a security hole in the booking system entirely on its own initiative. Links Nils Rollshausen's Apple Watch trust research Richard's blog post on Copilot fleet mode GitHub's fleet mode announcement Techerati on Apple's renewed UK encryption fight Original Penn State study, Mind Your Tone Follow up Penn State study, Does Tone Alter LLM Performance The Register on the Australian gym booking incident The Next Web on the Florian Roth pushback and liability question Socials X/Twitter Bluesky LinkedIn Facebook Threads Music Null Invocation, Monochrome Pulse