6 min read · 12 sources

Salesforce thinks it has the answer to the question every enterprise is about to hit: how do you give AI agents real access to your data without handing them the keys to everything? The Enterprise AI Harness is a six-capability architecture - context, agency, action, governance, security, and models - plus a new AI Control Plane for managing agents across the org. It’s open and composable, works with third-party models, and can be adopted in pieces or all at once.

That’s the good news. The rest of today’s newsletter is a reminder that the AI platform race is colliding with the messy reality of operating systems, identity, and files that weren’t built for agents. Microsoft confirmed a Windows update that kills audio with no fix. Sierra published field notes showing the MCP gateway problem is auth, not protocol. And macOS quietly broke a Keychain migration path that IT admins have relied on for years.

The MCP spec leaves gateway behavior undefined, so every enterprise invents its own auth, and that’s where the real cost lives.

Salesforce's Harness: six capabilities, one control plane

Source: salesforce.com ↗

The Enterprise AI Harness is Salesforce’s attempt to solve the problem that’s been nagging every AI platform team: reasoning models are great at generating text and terrible at knowing which database they’re allowed to query. The harness wraps agents in six named capabilities - context, agency, action, governance, security, and models - and ties them together with an AI Control Plane that manages agents the way a service mesh manages microservices.

The architecture is deliberately model-agnostic. You can run it with Salesforce’s models, third-party models, or a mix, and it plugs into external systems and agents. The pitch is that you don’t build the policy, audit, and security rails yourself - you buy them. For engineers, the interesting part is the Control Plane: it’s the piece that gives you a single view of what every agent is doing, which tools it can touch, and what it’s allowed to change. That’s the layer most enterprises are currently duct-taping together from Okta policies, API gateways, and hope.

Windows 11's audio-killing update has no fix yet

Source: windowslatest.com ↗

Microsoft confirmed that September’s Windows 11 update KB5124008 (Build 26200.9445 for 25H2, 26100.9445 for 24H2) can break USB Audio Class 1.0 devices. Symptoms include devices that fail to start, no sound at all, and an unresponsive Sound settings page. Device Manager shows a Code 10 error. The root cause sits in Windows Update itself, which means uninstalling the driver or disabling the device doesn’t fix it.

The audio issue is the headline, but the same update carries a stack of acknowledged bugs: Explorer.exe crashes, broken File History backups, Linux-based VM problems, Remote Desktop failures, and GPU issues on some AMD systems. Some users report that switching to two-channel audio helps with multichannel failures, but there’s no general fix pending. If you’re running 24H2, 25H2, or 26H1, this is the update to hold until Microsoft ships a patch.

The MCP gateway's real problem is auth

Source: workos.com ↗

Sierra’s Mihai Parparita published field notes on building an internal MCP gateway, and the headline is blunt: auth is the hard part, not the protocol. In their setup, 89% of employees connect agents to 45 services. The challenges are identity, per-tool scopes, consent, and audit - problems that don’t transfer cleanly from traditional enterprise software, where you authenticate a person once and they get a fixed set of permissions.

The MCP spec deliberately leaves gateway behavior undefined, so every enterprise invents its own. That’s the expensive work. If you’re building an agent gateway, read this before writing code: the integration layer is the easy 20%, and the auth layer is the 80% that will consume your team.

Security through obscurity just died

Source: theregister.com ↗

The Register’s argument is that AI killed security through obscurity, and it’s hard to disagree. Obscure software and unfamiliar legacy systems used to provide implicit protection - attackers had to know they existed and how they worked. AI removes that barrier by making discovery and exploitation cheaper.

The bigger point is remediation capacity. Finding more vulnerabilities accomplishes nothing unless teams can prioritize fixes, validate patches, and stop the same weaknesses from recurring. AI widens the attack surface faster than most orgs can patch it. The bottleneck isn’t detection anymore; it’s the queue of things you know about and can’t fix.

Dell's 52-inch monitor is a glorious mistake

Source: theregister.com ↗

Dell’s UltraSharp 52 is pitched as a replacement for four smaller monitors, cutting cable and stand clutter. The Register’s hands-on review found it genuinely impressive for visual work and multi-feed setups - and genuinely impractical for ordinary office work, with connection problems, awkward port access, and a footprint that’s hard to justify. It’s a fun toy. It is not a sensible purchase for most teams.

OpenAI's Agents API goes to public beta

Source: openai.com ↗

OpenAI’s managed Agents API is now in public beta. The service hosts the underlying harness and lets developers run agents in OpenAI-managed sandboxes, on their own infrastructure, or through supported providers. The pitch is reduced infrastructure management and engineering complexity - you bring the agent logic, OpenAI handles the harness. For enterprises that don’t want to build an agent runtime from scratch, this is a meaningful shortcut. The tradeoff is that you’re committing to OpenAI’s runtime and its governance model.

Boomi's September release adds Claude agent governance

Source: boomi.com ↗

Boomi’s September 2026 release makes Orchestrate generally available and adds Anthropic Claude Managed Agents as a native provider in Agent Control Tower, joining Boomi Agent Garden, Amazon Bedrock, and Snowflake Cortex. Agent details, versions, models, and token usage sync automatically. The new piece is governance: administrators can inspect agent session logs across users, and new connectors cover Zoom, Ping Identity, and other business apps. If you’re running multi-provider AI automation inside Boomi, this is the release that gives you a single pane of glass for agent behavior.

Slack's blurry images were a file-name bug

Source: slack-status.com ↗

Slack’s weekend incident made older image uploads appear blurry or fail to download across desktop, web, and mobile. The initial theory was a thumbnail-sharpening change, and rolling that back didn’t fix it. The real cause was a code change that caused older file records to lose their stored file names, so the file-serving system returned near-empty responses instead of full image data. New uploads were unaffected. A fix restored access, and a full investigation is underway. The lesson: file-serving bugs can look exactly like rendering bugs, and the fix path is different.

Okta's take: agents are new hires

Source: newsletter.getdx.com ↗

Okta Chief Architect Robert Lucero’s podcast on agentic AI and identity makes the case that identity systems are a control plane, not an orchestration layer. Treat AI agents like new hires: start them in sandboxes with tightly scoped access and just-in-time permissions. Agent identity definitions are still evolving - agents may operate under orchestrators, service accounts, or human owners - and non-determinism makes fine-grained permissions and visibility critical. Strong testing, CI, and review processes matter even more as AI generates more code.

macOS Tahoe 26.4 breaks Keychain migration

Source: appleinsider.com ↗

macOS Tahoe 26.4 introduced a change that prevents manually copying your login Keychain to another Mac. The Secure Enclave now guards the metadata key required to decrypt it, so the Keychain only works on the original Mac with its specific Secure Enclave. Testing confirms copying works on 26.3 and fails on 26.4. For IT teams, this is a ticking time bomb: users discover the failure only when they try to recover a Keychain on a new machine after a hardware failure. Plan your migration path now, before someone loses credentials in a recovery scenario.

iOS 27 lands Monday

Source: macrumors.com ↗

iOS 27 rolls out September 14, likely at 10:00 a.m. Pacific Time, alongside macOS Golden Gate and the rest of the OS 27 family. The update brings more advanced Siri AI capabilities. Expect time zone variations - 6:00 p.m. CEST in Europe, 5:00 a.m. NZST on Tuesday in New Zealand - and note that server delays are possible, so exact availability may vary.