We use both. Our customers run Microsoft 365 Copilot, and we run our own agents on Claude. So, we get asked a lot: which one is safer for the enterprise?
Start with a distinction that most comparisons blur. A model is the engine. An app (some people call it a harness) is everything built around the engine: the chat window, the connectors, the agents, the guardrails. Copilot is an app. It runs on models from OpenAI and Anthropic. Claude is a family of models, and Claude chat, the desktop and mobile apps, Cowork and Claude in Chrome are Anthropic’s own apps built around them.
That distinction matters because the two are strong at different layers. Anthropic builds some of the most secure models on the market. Microsoft wraps Copilot in the enterprise controls you already run: Entra, Purview, Defender and Intune. If your content lives in Microsoft 365, the second layer is the one you’ll feel.
|
The short version • Models and apps are different things. Industry benchmarks test models and the labs that build them. Anthropic leads there. • Copilot is an app that inherits your Microsoft 365 permissions, labels, DLP, threat protection and device rules. Claude’s own apps have controls of their own, but they don’t plug into those. • That gap is typical of standalone AI apps. ChatGPT Enterprise has the same one. • Claude’s models are also available inside Copilot. That gives you Microsoft’s controls, but you get Copilot’s app, not Claude’s. |
What the benchmarks measure
When you read that one AI vendor “ranks first for security,” ask which layer was measured.

Each kind of measure looks at a different layer. Anthropic leads on the model, Microsoft on enterprise controls.
The model. Prompt-injection tests put each model in the same test setup and try to make it follow hidden instructions. They measure the model on its own, not the product around it. In the largest public test so far (published March 2026, with its data shared with the US and UK AI safety institutes), all 13 frontier models proved vulnerable, and a Claude model held up best. Safety ratings such as the Future of Life Institute’s AI Safety Index grade the labs that build models (Anthropic got the top grade). Microsoft isn’t rated, because Copilot uses other companies’ models.
The app around the model. No public league table ranks AI apps on security. What exists are audits of the vendor’s own controls. Microsoft and Anthropic both hold SOC 2, ISO 27001 and ISO 42001, the AI management standard.
Your enterprise controls. Nobody benchmarks this, because it’s your tenant: your labels, your DLP rules, your threat protection, your managed devices. This is where Copilot’s position inside Microsoft 365 pays off, and the rest of this article is about it.
So, both statements are true. Anthropic’s models resist attacks better than most. Copilot sits inside a stronger set of enterprise controls. And a strong model doesn’t replace controls like DLP, any more than strong controls make up for a weak model.
No one is immune. Every major vendor fixed a serious prompt-injection flaw in the last two years: EchoLeak in Microsoft 365 Copilot (2025), the “Gemini Trifecta” in Google’s tools (2025) and a claude.ai data-exfiltration chain (2026). All were patched.
Where each one runs

Microsoft 365 Copilot sits inside your tenant, wrapped in the controls you already run.
Copilot keeps prompts, the data it retrieves and its responses inside the Microsoft 365 service boundary. Microsoft’s Data Protection Addendum and Product Terms apply, and Microsoft is the processor. One exception: when Copilot searches the web, it sends Bing a short generated query, which sits outside the DPA and the EU Data Boundary. An admin can turn web search off with one policy.

Claude’s apps are a separate service, protected by Anthropic’s own layers.
Claude’s apps run on Anthropic’s cloud. Claude Enterprise data is stored in the US, with a US-only processing option. Anthropic holds SOC 2 Type II, ISO 27001 and ISO 42001, and offers a HIPAA-ready setup with a BAA.
The direct comparison
Microsoft 365 Copilot against Claude’s own apps (chat, desktop, mobile, Cowork, Claude in Chrome) on the Enterprise plan, as of November 21, 2026.
|
Control |
Microsoft 365 Copilot |
Claude apps (Enterprise) |
|
Runs inside your tenant boundary |
Yes. Microsoft DPA and Product Terms |
No. Separate service under Anthropic’s terms |
|
Sign-in and Conditional Access |
Yes. Entra ID with full Conditional Access |
Partial. Entra SSO; the M365 connector can’t use location or sign-in frequency policies |
|
Model resistance to prompt injection |
Depends on the model. OpenAI or Anthropic models, plus Microsoft’s own injection classifiers |
Strong. Claude held up best in the March 2026 test, plus Anthropic’s classifiers |
|
Injection defense before content reaches a user |
Yes. Defender scans inbound mail; spam kept out of grounding; Safe Links |
No. Defenses run inside Claude, not upstream |
|
Perimeter defense (links, attachments) |
Yes. Defender for Office 365 |
No. IP allowlist and tenant restrictions control who reaches Claude |
|
DLP and keyword protection |
Yes. Purview DLP blocks labeled files and prompts with sensitive information (some parts in preview) |
Through partners. Netskope or Proofpoint inline checks |
|
Sensitivity labels and encryption honored |
Yes |
No. Not supported by Microsoft’s Purview connector for Claude |
|
Limit which SharePoint sites the AI can reach |
Yes. Restricted Content Discovery, Restricted SharePoint Search |
No. The connector searches every site the user can open |
|
Separate environments for agents |
Yes. Power Platform environments, environment DLP, dev/test/prod |
Partial. Child organizations with hard data boundaries (beta) |
|
Device management |
Yes. Intune app protection on Windows, iOS and Android |
Partial. An Intune-managed iPhone and iPad app |
|
Find other AI agents on company computers |
Yes. Defender and Agent 365 |
No |
|
Agent registry and identity |
Yes. Agent 365, an Entra Agent ID for every agent |
Partial. Owners control connectors; custom roles |
|
Audit and eDiscovery |
Yes. Purview audit, eDiscovery, legal hold |
Partial. Audit logs, Compliance API; Purview eDiscovery through a connector (preview) |
|
Retention |
Yes. Purview retention policies and labels |
Partial. One custom setting, 30 days minimum |
|
Insider Risk, Communication Compliance |
Yes |
Partial. Through the Purview connector (preview) |
|
Information Barriers, Customer Lockbox, just-in-time admin |
Yes |
No equivalent found |
|
Data residency |
Yes. EU Data Boundary, Multi-Geo |
Partial. US storage, US-only processing option |
|
Customer-managed keys |
Yes. Customer Key |
Yes. For eligible customers, US regions |
|
Certifications |
SOC 2, ISO 27001, ISO 42001, FedRAMP High (GCC High) |
SOC 2, ISO 27001, ISO 42001, HIPAA BAA |
Most of Claude’s “No” and “Partial” rows are about Microsoft 365 integration, not weak security. A standalone app can’t inherit controls that live inside your tenant. Microsoft’s own Purview connector for ChatGPT Enterprise lists the same gaps: no sensitivity labels and no DLP.
Where Microsoft’s stack makes the difference for enterprises
|
IF YOUR CONTENT IS IN MICROSOFT 365 Labels and DLP travel with it |
|
With Purview, you can stop Copilot from using files with chosen sensitivity labels, and block prompts that contain sensitive information. Claude’s apps can’t read your labels, and Microsoft’s Purview connector for Claude (in preview) covers audit and eDiscovery, not labels or DLP. To block sensitive data going into Claude in real time, add a partner such as Netskope or Proofpoint. |
|
IF YOUR CONTENT IS IN MICROSOFT 365 Watch what the connector can reach |
|
Anthropic’s Microsoft 365 connector sees what the user sees: every SharePoint site and OneDrive file they can open, across the whole tenant. It can’t be narrowed to chosen sites, and Anthropic’s documentation doesn’t say whether it honors Restricted Content Discovery or Restricted SharePoint Search. A Global Administrator has to consent to it first. If you do, set “Assignment required” on both of its Entra apps so only chosen people can use it. |
|
IF YOUR CONTENT IS IN MICROSOFT 365 Threats stop at the door |
|
Defender for Office 365 checks links, attachments and even hidden prompt injections in inbound mail before a person or Copilot opens them. Claude’s model defenses are strong, but they run inside Claude. Nothing on Anthropic’s side screens what reaches your people first. |
|
IF YOUR CONTENT IS IN MICROSOFT 365 Your device rules already apply |
|
Copilot runs in Office apps that Intune already manages on Windows, iOS and Android. Anthropic offers one Intune-managed app, for iPhone and iPad. The desktop and Android apps have no Intune app protection yet. |
Two more differences apply wherever your content lives.
Environments for agents. In the Power Platform, agents run in environments with their own DLP rules, so a development agent can’t use a production agent’s connectors. Anthropic’s closest match is a parent organization with separate child organizations, in beta.
Finding other agents. Microsoft can find AI apps and local agents on company computers and block them through Defender, Agent 365 and Intune. Anthropic doesn’t scan for other agents.
|
Questions to ask before you choose • Which controls protect this data today, and does the AI app inherit them? • Can you block sensitive data going out, or only see it afterwards? • Who can use the Microsoft 365 connector, and what can they reach? • Which agents will act, as well as answer, and who approves what they do? |
A short note on Claude inside Copilot
Here the model/app line matters most. Claude inside Copilot means Claude’s models running in Copilot’s app. You get Microsoft’s controls and terms, and you get Copilot’s features. You don’t get Claude’s apps: no Cowork, no Claude in Chrome, none of the Claude chat experience. It works like Copilot with a different engine.
Anthropic has been a Microsoft subprocessor for Claude in Copilot since January 2026, announced in December 2025. Before that, Anthropic was an independent processor and Microsoft’s terms didn’t apply.
- Claude’s models are available in Researcher, Copilot Studio, Edit with Copilot in the Office apps, and Microsoft Copilot.
- Microsoft’s Product Terms and DPA cover them, and Microsoft’s controls apply as they do to the rest of Copilot.
- They’re on by default in commercial tenants and off by default in the EU, EFTA and UK. They aren’t available in GCC High or DoD.
- They sit outside the EU Data Boundary today. Microsoft has announced they move inside in early December 2026.
- Models that keep data (for example Fable 5) still run under Anthropic’s own terms and need an explicit admin opt-in.
Prices (US list, checked November 21, 2026)
|
Plan |
Price |
|
Microsoft 365 Copilot Business add-on |
$21 per user per month (promotional $18 on annual billing through December 31, 2026) |
|
Claude Team, standard seat |
$20 per user per month on annual billing, $25 monthly |
|
Claude Enterprise |
$20 per seat per month on annual billing, minimum 20 seats, plus usage at API rates |
Prices change often, so confirm them before you budget.
Which of these controls would you need before letting an AI app near your data?
Sources (checked November 21, 2026)
- Future of Life Institute: AI Safety Index, Summer 2026 (Jul 2026)
- arXiv: Indirect prompt injection arena results (Mar 2026) and NIST CAISI: Insights from a large-scale AI agent red-teaming competition (Mar 23, 2026)
- Microsoft Learn: Copilot and AI subprocessors (Anthropic) (Nov 20, 2026)
- Microsoft 365 Message Center: MC1546438, Claude in the EU Data Boundary (Nov 10, 2026)
- Microsoft Learn: Prompt defense in depth (updated Nov 3, 2026)
- Microsoft Learn: DLP for Microsoft 365 Copilot
- Microsoft Learn: Purview for Claude Enterprise and Purview for ChatGPT Enterprise
- Microsoft Security blog: Agent 365 generally available (May 1, 2026)
- Anthropic: Microsoft 365 connector security guide
- Anthropic: Use Claude Cowork safely and Use Claude in Chrome safely
- Anthropic: Certifications, IP allowlist and Claude for Intune
- Anthropic: Compliance API security partners (May 21, 2026)
- Microsoft and Anthropic pricing pages
