Daily AI brief · May 26, 2026
Anthropic Puts Claude Agents Inside Your Firewall
Google's Gemini 3.5 Flash rewrites the "cheap model" playbook, Anthropic lets Claude agents run inside your own perimeter, Microsoft's AI Builder credit clock is ticking toward November, and a new WatchGuard study hands MSPs a concrete sales number to carry into every security conversation.
1. Anthropic Puts Claude Agents Inside Your Firewall *(Anthropic / The Decoder / InfoQ)*
Anthropic has expanded its Claude Managed Agents platform with two enterprise-focused capabilities: self-hosted sandboxes and MCP tunnels — directly addressing the recurring challenge where organizations want to use autonomous agents but cannot allow execution environments or internal systems to leave their security perimeter.
Self-hosted sandboxes let companies run agent tool calls on their own infrastructure or through managed providers like Cloudflare, Daytona, Modal, and Vercel. MCP tunnels connect agents to MCP servers on a private network through a single outbound, end-to-end-encrypted connection — no inbound firewall rules required — letting agents tap internal databases, private APIs, or ticketing systems as tools.
Self-hosted sandboxes are in public beta; MCP tunnels are research preview and require access requests.
What it means for SMBs / MSPs: If a client has been stalling on an AI agent pilot because their compliance team won't let data leave the building, this removes the most common blocking objection — bring it back to the table now.
2. Google Gemini 3.5 Flash Goes GA — Faster Than Last Year's Pro, at 3× the Flash Price *(Google I/O 2026 / LLM Stats / Simon Willison)*
Gemini 3.5 Flash went GA on May 19, 2026. It outperforms Gemini 3.1 Pro on the coding and agentic benchmark suite, runs 4× faster than comparable frontier models, and ships at $1.50/$9 per 1M tokens with a 1M-token context window.
The catch:
it comes with a notable price bump — the new 3.5 Flash is 3× the price of the Gemini 3 Flash Preview it replaces.
If you are routing the same workload that ran on 3 Flash to 3.5 Flash, plan for somewhere between 3× and 5× on the bill , since the model also spends more output tokens per agentic turn.
What it means for SMBs / MSPs: Any client or internal tool currently hard-coded to the old Flash endpoint needs a cost-model review before auto-upgrading — the capability jump may be worth it, but the bill jump will surprise unprepared teams.
3. Microsoft's AI Builder Credit Sunset: Five Months and Counting *(Microsoft Learn / Microsoft Docs)*
After November 1, 2026, seeded AI Builder credits will be removed for all new and existing customers, including those with an Enterprise Agreement. Copilot Studio Credits will be required for continued access.
There is no automatic conversion of old AI Builder credits to Copilot Credits
— they are separate currencies.
Organizations that fail to plan will experience service disruption on November 1, 2026.
What it means for SMBs / MSPs: Any Power Automate or Power Apps flow using AI Builder features needs a credit audit now — run the usage report in Power Platform admin center, map it to the new Copilot Credits rate card, and budget the difference before Q4 renewals lock in.
4. WatchGuard Study: 91% of SMBs Fear AI-Driven Attacks, 54% Can't Do 24/7 Monitoring *(WatchGuard / GlobeNewswire — May 20, 2026)*
According to a study of nearly 1,000 IT and cybersecurity leaders across 20 countries, 91% of organizations are concerned about AI-driven cyberattacks, while 75% experienced at least one cyber incident in the past year. At the same time, 54% say they lack the ability to deliver continuous 24/7 monitoring and response, and 67% need additional support to meet growing compliance demands.
The report finds that 44% of organizations are willing to pay more for AI-powered detection and response, highlighting a growing expectation that security providers deliver not just protection, but measurable outcomes.
What it means for SMBs / MSPs: The WatchGuard data gives you third-party validation for a managed security services conversation — lead with the 91%/54% figures in your next QBR or prospect call.
5. Anthropic Launches "Claude for Small Business" — Pre-Built Workflows for the Tools SMBs Already Use *(CX Today / Guardz — May 14, 2026)*
Anthropic launched Claude for Small Business, a new AI offering designed to help small and medium-sized businesses adopt AI through the software they already use every day. The launch aims to close a long-standing adoption gap in the SMB market, where many business owners lack the time, technical expertise, or internal resources to experiment with enterprise-grade AI tools.
The release includes software connectors and pre-built workflows that allow Claude to work directly inside business tools many SMBs already use, including PayPal, HubSpot, Google Workspace, and Microsoft 365.
What it means for SMBs / MSPs: This is a low-friction entry point for clients who've resisted AI because it felt too technical — the activation flow is connector-first, not model-first.
Claude for Small Business — pre-built agentic workflows inside HubSpot, Google Workspace, and M365
The offering works through a simple activation inside Claude Cowork, allowing users to connect their existing apps, choose a specific workflow, and let Claude complete tasks while requiring human approval before anything is sent, posted, or paid.
- Who it's for: AI champions at SMBs (5–100 employees) who've been told "AI is too complicated" — or MSPs looking for a demo-ready product for hesitant clients.
- How long a first test takes: 30–45 minutes to connect one app (HubSpot or Google Workspace) and run one pre-built workflow end-to-end.
- First step: Go to claude.ai, navigate to the Cowork section, connect your HubSpot or Google Workspace account, and activate one workflow — customer follow-up drafts are the fastest win.
Check whether anyone watches for attacks around the clock, not just during business hours. The WatchGuard study found 91% of businesses worry about AI-driven attacks and 54% can't provide 24/7 monitoring. AI helps attackers move faster than a one-person IT team can respond. Compare what always-on monitoring costs with what a breach would cost you.
*For MSPs:* For clients with one internal IT person or none, the WatchGuard stats frame always-on, AI-assisted monitoring against the cost of a breach.
Don't assume Gemini 3.5 Flash is a drop-in cost-neutral upgrade from Gemini 3 Flash.
If you are routing the same workload that ran on Gemini 3 Flash to 3.5 Flash, plan for somewhere between 3× and 5× on the bill
— the model is more capable but also more verbose per agentic turn. Audit token usage *before* flipping the model ID in any production flow or client deployment.
The "agent inside the perimeter" architecture is becoming the enterprise standard — and it's now trickling toward mid-market.
The release comes as enterprise AI tools increasingly move toward deployments that combine cloud-based intelligence with customer-controlled execution environments, reflecting a wider industry emphasis on security, compliance, and infrastructure flexibility.
What Anthropic shipped this month (self-hosted sandboxes + MCP tunnels), Microsoft has been building toward with Copilot Studio's VPC-adjacent controls, and Dell is pitching as "AI-native infrastructure" to the channel. The pattern is clear: the question is no longer *whether* AI agents touch internal systems — it's *who controls the execution environment when they do*. MSPs who can speak fluently to data residency, audit logging, and runtime boundaries will win mid-market AI deals that pure SaaS vendors can't close. Start building that conversation now.