If I only use Claude for Excel and never install the PowerPoint add-in, am I completely free of this risk?
Using Claude for Excel alone, without also chaining the PowerPoint add-in, does mean this specific cross-app data flow risk doesn't apply — the scenario Anthropic's documentation describes assumes both add-ins are connected to the same Cowork session, which is the precondition for data having a path to flow "from one app into another" in the first place. If your workflow stays entirely within Excel, that specific risk path simply doesn't exist.
That doesn't mean there's nothing else to watch for, though — even with a single app, the general risks of Cowork accessing files and folders (overly broad granted scope, sensitive data being involved) still apply; it's specifically "content getting carried into a second app you weren't expecting" that won't happen. If you later decide to add the PowerPoint add-in and chain a fuller workflow, that's the point at which this particular risk needs reassessing.
If I genuinely need to work with an Excel file that contains sensitive tabs, is there another option besides saving a clean copy?
Saving a clean version is the most direct and safest option in Anthropic's own recommendation, but if that's not feasible (say, formulas link the sensitive tab to the tab you actually need, and splitting them would break the calculation logic), another workable middle ground is altering the fields in the sensitive tab into a form that doesn't disrupt the formulas but strips out the actually sensitive content — swapping real names for codes, or actual dollar amounts for ranges — ensuring what Claude reads doesn't include the specific information that genuinely needs protecting, while keeping the file's structural integrity intact.
If even that isn't feasible, the more solid fallback is avoiding cross-app context sharing for this particular task entirely — falling back to working within a single app and manually carrying content over yourself at each step. That sacrifices the efficiency automation buys you, but in exchange, the scope of data flow stays entirely under your own control, with no unexpected content getting swept along.
Once Claude has carried content that shouldn't be there into a deck, is there truly no way to catch it before I happen to notice on my own?
There's actually still room to catch it, and the key factor is how quickly. Anthropic's recommended approach is, before the deck actually goes out (shared with a client, uploaded to a shared folder, sent by email), to explicitly ask Claude to list the specific source of every figure and chart in the deck. That step essentially lays the data flow trail out for you to inspect, rather than passively waiting for you to spot an anomaly on your own. Do this step before sending, and the vast majority of content that shouldn't be there gets caught right at this stage.
What's genuinely hard to remedy is discovering the problem after the deck has already gone out — at that point it's moved beyond what Claude Cowork as a tool can handle, and become a post-disclosure response problem instead (contacting the recipient to request deletion, assessing whether it rises to a reportable data breach). That's also why Anthropic's recommendation focuses on the "verify sources before sending" step — it's the point in the whole process with the lowest remediation cost, and the last moment where the problem can still be caught in time.
Besides Excel and PowerPoint, do other Office add-ins like Claude for Word or Claude for Outlook follow the same risk logic?
The same underlying logic applies, though the actual risk profile varies by the nature of the app. The core principle stays the same: whenever multiple Office add-ins are connected to the same Cowork session, context gets shared between them, and data has a path to flow from one app to another. Take Outlook as an example — if it's chained into the same session as Excel and PowerPoint, email content and contact information in the inbox could theoretically also become context Claude draws on while handling a task in another app, extending the risk scenario from "Excel data leaking into a presentation" to "inbox content leaking into a presentation or document."
The practical judgment call is the same: before starting a task involving multiple Office add-ins, take stock of every app that'll be connected within that session and what content each might contain that shouldn't leak out — rather than being careful specifically about the Excel-PowerPoint combination while overlooking other sensitive information that might be sitting in an inbox or a document. Any app connected within the same session falls under the same logic: the reach of shared context is larger than you might assume.
Claude for Excel and Claude for PowerPoint, when used together inside Claude Cowork, share the full context of a single session — meaning you can ask Claude to read revenue figures in Excel and drop a chart straight into a PowerPoint deck, with no manual copy-pasting and no need to re-explain what you're working on. This cross-app workflow genuinely eliminates a lot of repetitive work, but there's a line in Anthropic's own safety guidance worth reading closely first: Claude might analyze data in Excel and move a chart into a presentation — without you explicitly directing that transfer.
Once you've installed both the Claude for Excel and Claude for PowerPoint add-ins and use them inside Cowork, the two apps connect to the same Cowork session — meaning content Claude has read and analyzed in one app becomes context it can draw on while working in the other. Anthropic's documentation describes the exact scenario: Claude might first analyze data in Excel, then move a chart into a presentation, and that entire transfer may not be something you directed step by step — it's Claude deciding, based on its understanding of the task's context, what to carry over on its own.
The reason this feature exists is straightforward: if you ask Claude to "build a deck using the Q4 numbers in Excel," you obviously want it to read the correct figures and pick the right chart itself, rather than requiring you to manually copy-paste every data point in advance before it can even start. Sharing context across apps is exactly what makes this kind of "describe the whole task in one sentence" workflow possible. The issue isn't that the feature is badly designed — it's that if you're not aware data flows automatically, you might unknowingly end up carrying information into a presentation that shouldn't have been there.
Picture a common setup: alongside the Q4 revenue figures meant for the deck, the same Excel workbook might have other tabs containing employee salary details, unannounced M&A valuation numbers, or a client contact list with personal information. If you ask Claude to "analyze this Excel file and build a deck for the client" without scoping it precisely, Claude may — based on its understanding of the task — carry content it judges relevant into the draft deck. Not because it deliberately meant to leak anything, but because it read the context of the entire workbook, and you only checked afterward whether the deck looked complete, without individually confirming that every chart and every figure's source matched the range you actually intended.
Anthropic's safety guidance is fairly concrete on this: be aware that data may flow from one application into another during a single Cowork session, and avoid working with sensitive information in these add-ins while Cowork is active. The logic behind that is simple — rather than checking after the fact whether every transferred piece of content is appropriate, it's better to reduce, at the source, how much sensitive content ever ends up within reach of the same session's context. If a given task genuinely needs to work with a workbook containing sensitive tabs, the safer move is saving the tabs you actually need into a clean, separate file and handing only that clean version to Cowork, rather than opening up access to the entire original file.
If you're only using Claude for Excel by itself on a single spreadsheet, without also chaining PowerPoint, the risk scope is relatively contained — content won't get carried into a second app you weren't expecting. The cross-app risk fundamentally comes from the fact that the context's reach extends further than whatever single file you happen to be looking at right now — which is also why Anthropic's documentation discusses this alongside the principle of local folder access scope: both share the same underlying logic, that the scope you've granted determines what Claude can theoretically reach, and cross-app context sharing simply expands that scope from "one file" to "every app you've connected within this session."
If you're planning to chain Excel and PowerPoint through Claude Cowork to build a deck meant for an external audience — clients, investors, another department — spend a minute upfront confirming whether the Excel workbook you're about to reference mixes in other tabs or data that shouldn't appear in that external deck. If the answer is yes, saving the relevant portion into a standalone file first is more reliable than reviewing the deck slide-by-slide afterward. It's also worth building a habit once the output is ready: before sending the deck out, explicitly ask Claude to list which Excel tab and cell range each figure and chart in the deck came from. Having sources spelled out one by one catches whether data got pulled in from somewhere it shouldn't have far more reliably than just eyeballing whether the layout and text read smoothly.