Daily AI brief · May 8, 2026
Anthropic Drops 10 Finance Agents, Claude Hits M365
Anthropic launches pre-built financial agents that plug into Excel, Word, and Outlook; OpenAI rolls GPT-5.5 as the new ChatGPT default; CISA and allies issue a stark agentic AI security warning; and self-hosted AI infrastructure is turning into an open wound for SMBs.
1. Anthropic Drops 10 Pre-Built Finance Agents — and Claude Now Lives in Excel, Word & Outlook *(Bloomberg / Fortune / Anthropic)*
Anthropic unveiled a set of new AI agents designed to handle a broader mix of financial services tasks — the Claude maker is releasing tools that can draft pitch decks for client meetings, review financial statements, and escalate cases for compliance review. The 10 new tools are aimed at professionals across banking, insurance, asset management, and financial technology.
Critically for non-Wall Street shops:
Claude now works across Microsoft Excel, PowerPoint, Word, and Outlook through Microsoft 365 add-ins, and context moves between applications automatically, so work that starts in a spreadsheet can continue in a deck without re-explaining everything.
What it means for SMBs / MSPs: Even if your clients aren't hedge funds, these agent templates are reference architectures — the pattern of skills + connectors + subagents is exactly the playbook for building your own vertical agents for accounting, insurance, or operations clients.
2. GPT-5.5 Is Now the Default ChatGPT Model — With Memory and Spreadsheet Sidebars *(OpenAI Release Notes)*
GPT-5.5 Instant is rolling out to all ChatGPT users, replacing GPT-5.3 Instant as the default model and in the API as `chat-latest`; paid users retain access to GPT-5.3 for three months before it's retired.
Beyond the model swap:
ChatGPT Business now supports ChatGPT for Excel and Google Sheets worldwide, bringing a spreadsheet-native sidebar for building, cleaning, updating, and explaining workbooks, with a free preview for Business customers through June 2, 2026.
What it means for SMBs / MSPs: The free Business preview of ChatGPT in Excel is your lowest-friction demo this month — hand it to a skeptical owner with a messy spreadsheet and let the tool close itself.
3. Microsoft Copilot Cowork Adds Mobile Delegation and Deeper Power Platform Hooks *(Microsoft 365 Blog)*
The shift behind Copilot Cowork is moving beyond chat and into execution — helping users delegate real tasks and have them completed. Microsoft is now introducing Cowork on iOS and Android so users can delegate work the moment they think of it, on a commute or between meetings, and come back to a finished outcome.
On the automation side,
the latest Power Automate updates introduce agent-based features for authoring, optimization, and self-healing that apply to desktop flows and Copilot Studio-driven actions in cloud flows, and Copilot Studio has added new capabilities for customizing agents built with the Agent Builder in Microsoft 365 Copilot.
What it means for SMBs / MSPs: Self-healing desktop flows are the quiet headline here — automations that fix themselves when a UI changes reduce your support burden on RPA deployments significantly.
4. CISA, NSA, and Allied Agencies Issue Joint Agentic AI Security Guidance *(ASIS / CISA)*
New guidance from CISA, NSA, the UK NCSC, Australia's ACSC, and other allied agencies addresses key cybersecurity challenges associated with introducing agentic AI into IT environments, strongly recommending that organizations align agentic AI risks with their existing security postures and restrict agentic AI systems' access to sensitive data and critical systems.
The specific concern:
agentic AI is vulnerable to many of the same risks affecting other LLM-based systems, such as prompt injection attacks — and malicious actors do not need to learn new skills to exploit them, but the complexity inherent to AI systems makes securing them more challenging than traditional software.
What it means for SMBs / MSPs: This is now a compliance conversation, not just a technical one — clients deploying any autonomous agent (Copilot Studio, n8n, Make, etc.) need a documented access-control and approval-flow policy before something acts on their behalf.
5. Exposed Self-Hosted AI Infrastructure Is a Ticking Clock *(The Hacker News)*
While the software industry has made genuine strides on security, the furious pace of AI adoption is putting that progress at risk — businesses are moving fast to self-host LLM infrastructure, drawn by the promise of AI as a force multiplier, but speed is coming at the expense of security.
A scan of over 1 million exposed AI services found
repeated insecure patterns: insecure defaults, misconfigured Docker setups, hardcoded credentials, and applications running as root with no authentication on fresh installs.
What it means for SMBs / MSPs: If you've helped a client spin up an Ollama instance or any self-hosted model server, run an auth and exposure check this week — this is already an active attack surface, not a future one.
ChatGPT for Excel (Business free preview, through June 2)
What it is: A spreadsheet-native AI sidebar inside Excel that can build formulas, clean data, explain workbook logic, and answer questions about the numbers — no pivot-table expertise required.
Who it's for: SMB owners, ops managers, or any client who lives in Excel but spends more time fixing it than using it. Also a dead-simple demo tool for MSPs and AI champions making the "AI is practical" case.
How long a first test takes: Under 30 minutes — open an existing business workbook, ask it to explain a formula, then ask it to build a summary row. That's the demo.
One concrete first step: In your ChatGPT Business admin portal, enable the Excel integration under Apps, then share the enable link with one power-user client contact. Let them try it on a real file before June 2 at zero cost.
If someone set up a self-hosted AI server like Ollama for you, confirm it requires a login. A scan of over 1 million exposed AI services found insecure defaults, hardcoded credentials and apps running with no authentication on fresh installs. It's a quick check, and it's an active attack surface now. Also make sure it isn't reachable from the internet.
*For MSPs:* Community banks, RIAs and insurance agencies are a fit for Anthropic's finance agent templates inside the Microsoft 365 apps they already use.
Don't spin up a self-hosted Ollama or open-source model server for a client without a hardened access-control checklist. A fresh scan of 1 million exposed AI services found auth-free installs and hardcoded credentials are the norm, not the exception — you'll own the liability when it becomes a breach.
Agentic AI is officially a governance and liability category, not just a capability one. Between the joint CISA/NSA/NCSC guidance, Microsoft's agent approval-threshold controls in Power Apps, and Anthropic's explicit "human in the loop" language for its finance agents, the industry is converging on a model where every autonomous action must be traceable, auditable, and reversible. For MSPs, this is the wedge for a new recurring service line: AI governance audits that assess what agents have permission to do, what data they can touch, and who gets alerted when they act. Clients deploying Copilot Studio, n8n, or any multi-step automation today will need that paper trail tomorrow — likely sooner than they expect.