6 min read · 12 sources

Three of the most-used AI services went dark at the same time on Thursday, and a chunk of the Windows-on-ARM workforce couldn’t open their email or chat while it was happening. ChatGPT was out from 11am to 12:55pm ET, Claude stayed down for around four hours starting at 7:37am ET, and Grok was unreachable for roughly 3.5 hours. The overlaps hit search, file uploads, agents, voice, image generation, the APIs, and the Codex CLI. It happened to be launch day for OpenAI’s GPT-6 Astra, which the company is positioning as an AGI approximation. Analyst Carmi Levy put it bluntly: enterprises are “uncomfortably exposed” because they keep automating faster than they plan for AI failures.

Claude was down for about four hours, Grok for roughly 3.5, and ChatGPT from 11am to 12:55pm ET, all on the same day OpenAI launched GPT-6 Astra

Microsoft Teams and New Outlook crash on ARM-based Windows 11

Source: bleepingcomputer.com ↗

Microsoft is working on a bug that breaks Teams and the new Outlook on ARM-based Windows devices after the August cumulative updates. Surface Laptop 7 and Surface Pro 11 are the loudest victims, particularly freshly imaged machines that haven’t pulled Store updates yet. Classic Outlook, Word, and Excel are unaffected. The temporary fix is updating the Auto Super Resolution Package to 1.0.19.0 or later through the Microsoft Store. The root cause hasn’t been disclosed, and this sits on top of a string of August update issues around printing, PDF export, and RGB peripherals. Admins on ARM fleets should patch the ASR package now and stop imaging new devices until the permanent fix lands.

Salesforce bundles Slack, Tableau Next, Agentforce AI, and security into three SKUs

Source: salesforce.com ↗

Salesforce is collapsing its catalogue of add-ons into Core, Advanced, and Max editions across Agentforce Sales, Service, and Industries. Each tier bundles Agentforce AI, Slack and Slackbot, Tableau Next, enterprise data security, and a Premier Success Plan. Max delivers 60% more value than the previous Agentforce 1 Edition at no additional cost, and existing Agentforce 1 customers can upgrade for free. Credit allocations scale from 500,000 Flex Credits on Core, to 1 million on Advanced, to 2.75 million on Max. Salesforce also said future Max editions will include Headless 360 capacity, which pushes the platform toward agent and app callers rather than only human seats. For anyone tracking per-seat versus per-outcome pricing, this is Salesforce moving further away from seats.

77% of enterprises re-evaluate AI vendors every six months

Source: techcrunch.com ↗

Madrona surveyed 150 enterprise IT leaders and found 74% plan to expand AI budgets, but fewer than half of AI pilots reach full production. The sharper finding is that 77% re-evaluate AI vendors every six months or continuously, a “fast in, fast out” model that turns even a signed enterprise deal into a renewable, fragile revenue line. Separate Andreessen Horowitz research found 50% of technical AI buyers want pricing tied to work outcomes instead of tokens. Together, the two studies describe a market where AI ARR is structurally less durable than traditional SaaS, and where the unit economics are being rewritten in real time.

Data sovereignty is now an architecture problem, not a compliance one

Source: cio.com ↗

The IBM 2025 Cost of a Data Breach Report puts the global average breach at $4.9 million, which is why governance now lives at the board level. The piece argues that knowing which region hosts your data is no longer enough. Multi-cloud, AI workloads, and vendor churn mean sovereignty has to be expressed in portability and operational independence, not just geography. Engineers building cross-region AI pipelines should treat vendor concentration as a first-class risk and design for data egress and key control from day one, because the next regulator or geopolitics event will not wait for a refactor.

Okta ships Auth0 tools for AI agents and B2B logins

Source: itbrief.com.au ↗

Okta’s Auth0 group released AI Identity for Agentic Commerce, which verifies that an AI shopping assistant is acting on behalf of a real, authorized customer before it checks out. The launch also includes tools for companies building their own AI shopping agents, B2B Connect for adding enterprise SSO and federation without ripping out existing identity infrastructure, and an expanded tenancy management service for multi-tenant platforms. For anyone wiring agents into commerce flows, the unsolved problem is now squarely on the table: how do you prove a non-human caller is allowed to spend money or pull data, and how do you revoke that trust when it goes sideways.

JFrog turns its AI Catalog into an agent control plane

Source: jfrog.com ↗

JFrog’s AI Catalog started as a model registry and is now positioned as an AI control plane spanning models, skills, plugins, and MCP servers, all treated as native artifacts inside the software supply chain. The framing matters: agents don’t just load a model anymore, they assemble components from many open sources at machine speed, which sidesteps the traditional security review. JFrog’s answer is “agent immunization”: govern which AI assets an agent is allowed to consume, scan for shadow AI, and enforce policy at runtime via Agent Guard. Anyone who used to vet a single model once and call it done needs a new mental model, because the attack surface is now whatever the agent pulls mid-execution.

The human in the loop is often human-adjacent-to-the-loop

Source: cio.com ↗

A growing pile of governance write-ups makes the same point: many “human in the loop” setups give overseers the ability to see or flag a decision, but not to stop or override it. Cloudflare’s Doug Shepherd calls the pattern performative governance, because it gets projects through review without doing the actual risk work. The fix is mechanical: overseers need the domain knowledge to act when the AI is not in the loop, and the authority to actually intervene. Treat it like any other operational control, with drills, not slideware.

NASA's People Knowledge Graph

Source: memgraph.com ↗

NASA’s People Analytics team built a People Knowledge Graph using Memgraph on Docker and EC2, an on-prem Ollama LLM server, S3, and GQLAlchemy. The graph connects employees, skills, projects, and career paths, and exposes queries like “who has worked on similar AI projects across centers” plus a GraphRAG chatbot for natural-language skill discovery. The interesting part is the stack: graph traversal plus LLM-based skill extraction, all on secure internal AWS. For anyone running expert-find or staffing analytics, the same shape works: keep the model inside your perimeter, let it enrich a graph, query with Cypher or NL.

Rootly surfaces related incidents into the active Slack channel

Source: rootly.com ↗

Rootly’s on-call agent now posts up to three related past incidents into the active Slack incident channel, with relevance, severity, time to resolve, and links to the original. It runs without prompting and keeps the thread compact. The target problem is familiar: the person who fixed the last outage isn’t the person on call this time, and the institutional memory lives in postmortems nobody has time to read at 2am. Anyone running an incident process should be looking at this pattern, because cutting MTTR by eliminating the search step during an active page is free time.

Teams web is moving to teams.cloud.microsoft

Source: windowsreport.com ↗

Microsoft is redirecting Teams web users to teams.cloud.microsoft and has given admins until 2027 to make sure the new domain is allowed through proxies, firewalls, and other egress controls. If your network policy still blocks anything that isn’t teams.microsoft.com or *.office.com, web Teams is about to start failing for your users.