Daily AI brief · May 27, 2026
Anthropic Brings Claude Agents Inside Your Security Perimeter
Anthropic's self-hosted sandboxes and MCP tunnels finally let agents touch private data without exposing it; plus Google's Gemini 3.5 Flash goes GA, Microsoft's Insider Risk now watches AI agents, the Colorado AI Act is 34 days from enforcement, and Claude for Small Business hits its two-week mark.
1. Anthropic Opens the Perimeter: Self-Hosted Sandboxes + MCP Tunnels Go Live *(The New Stack / InfoQ / The Decoder)*
Anthropic has expanded its Claude Managed Agents platform with two enterprise-focused capabilities — self-hosted sandboxes and MCP tunnels — specifically to address the problem of organizations that want autonomous agents but cannot let execution environments or internal systems leave their security perimeter.
With self-hosted sandboxes, tool execution moves into the customer's own infrastructure; files and repositories never leave the company's environment, and existing network policies, audit logging, and security tools stay in place.
MCP tunnels connect agents to MCP servers on a private network without exposing them to the public internet — a lightweight gateway opens a single outbound encrypted connection, with no inbound firewall rules or public endpoints required.
Critically:
the agent loop itself does not move — orchestration, context management, and error handling still run on Anthropic's servers.
What it means for SMBs / MSPs: If you've had a client stall on deploying a Claude agent because their compliance team objected to data leaving the perimeter, this is the unlock — but read the fine print on what "self-hosted" actually covers before you oversell it.
2. Gemini 3.5 Flash Goes GA — Frontier Speed at Budget Pricing *(LLM Stats / Digital Applied)*
Gemini 3.5 Flash is now generally available, delivering what Google calls frontier-level intelligence at 4x the speed of comparable models, priced at $1.50/$9 per 1M tokens with a 1M token context window, and beating Gemini 3.1 Pro on coding and agent benchmarks.
What it means for SMBs / MSPs: At that price point and context window, this is a serious option for RAG-heavy document workflows (contracts, manuals, client records) where you want speed and don't need the most expensive model.
3. Microsoft Insider Risk Now Governs AI Agents *(MSFTNewsNow / LevelUpM365)*
Microsoft has extended Insider Risk Management to AI agents across Copilot Studio, Azure AI Foundry, and Agent 365 — admins can now define policies that detect, flag, or block risky AI agent activity, especially when sensitive data is accessed, exported, or combined in unusual ways. This acknowledges that AI agents can exfiltrate or misuse data just as quickly as human users.
Separately,
GPT-5.5 Reasoning is now rolling out across Microsoft 365 Copilot in Copilot Chat, Word, Excel, and PowerPoint.
What it means for SMBs / MSPs: If you're deploying Copilot Studio agents for clients, now is the time to configure IRM policies before an incident triggers them first — treat AI agents as you would privileged user accounts.
4. Claude for Small Business Hits Its Two-Week Mark *(CXToday / Guardz)*
Anthropic launched Claude for Small Business on May 13, 2026 — a new offering designed to help SMBs adopt AI through the software they already use, aiming to close a long-standing adoption gap where business owners lack the time, technical expertise, or internal resources to experiment with enterprise-grade AI tools.
The package includes software connectors and pre-built workflows that let Claude work inside tools like PayPal, HubSpot, Google Workspace, and Microsoft 365, with a human-approval gate before anything is sent, posted, or paid.
What it means for SMBs / MSPs: Two weeks in, this is worth a hands-on demo session with clients who use HubSpot or QuickBooks — the 15 pre-built workflows are a low-friction starting point for AI adoption conversations.
5. Colorado AI Act: 34 Days to Enforcement *(Wilson Sonsini / Kiteworks)*
The Colorado AI Act is slated to come into effect on June 30, 2026, placing substantial new responsibilities on AI developers and deployers — including requirements to avoid algorithmic discrimination, develop a risk management policy, implement notices, and conduct impact assessments.
Simultaneously,
cyber insurance carriers are introducing "AI Security Riders" that condition coverage on documented security practices, and organizations without robust AI risk management may face coverage denials or prohibitive premiums.
What it means for SMBs / MSPs: If any of your clients operate in Colorado or use AI for automated decision-making (hiring, pricing, credit), they need a documented AI risk policy in hand before July 1 — and their cyber insurer may ask for proof either way.
Claude Managed Agents — Self-Hosted Sandbox (Public Beta)
Self-hosted sandboxes, now in public beta, allow tool execution to run on infrastructure controlled by the customer or through managed providers such as Cloudflare, Daytona, Modal, and Vercel — while Anthropic manages orchestration, context handling, and recovery logic. The approach gives enterprises better control over network policies, audit logging, runtime configuration, and data residency.
- Who it's for: MSP techs and AI champions supporting clients in healthcare, legal, finance, or any vertical with a compliance-conscious IT or legal team.
- How long a first test takes: 60–90 minutes to spin up a sandbox using Modal or Vercel as the managed provider, connect it to a test repository, and run a simple file-summarization task.
- First step: Go to the Claude Console → Workspace Settings → Sandboxes, enable the public beta, and connect one of the day-one managed providers.
Setup is managed through workspace settings in the Claude Console by organization admins.
Start with a non-production repo.
Write down why you've said no to AI agents, then ask your IT provider if it still holds. For many law, accounting and HR offices, the reason is that client data can't leave their environment. Anthropic's new self-hosted sandboxes, now in beta, run agents on your own infrastructure. That may not change your answer, but check before you rule it out.
*For MSPs:* A short proof of concept on the client's own infrastructure gives compliance teams a credible answer on where code runs and where files land.
Positioning Anthropic self-hosted sandboxes as "fully on-premise Claude."
Self-hosted sandboxes are not a fully on-premise deployment. If a client's compliance requirement is that nothing touches external infrastructure at all, this does not fully solve that problem yet.
Selling it as complete data sovereignty will backfire the moment a security team digs into the architecture.
AI agent governance is becoming a product category, not a feature.
Microsoft Agent 365 provides a unified control plane for agents, enabling IT, security, and business teams to observe, govern, and secure agents across the organization — applying to any agents used, whether built on Microsoft AI platforms, delivered by ecosystem partners, or introduced through other means.
Anthropic is doing the same thing from the infrastructure side with sandboxes and MCP tunnels. The direction is clear: within 12 months, every serious enterprise AI platform will need to answer the question "where does this agent execute, what can it touch, and who gets the audit log?" SMBs that get ahead of this now — by treating agents as managed identities rather than chat tools — will have a material compliance advantage when regulators and insurers start asking the same question.