Data Security and Privacy: Can You Trust It with Company Files

If you want WorkBuddy to handle contracts, reports, or client lists, the first hurdle is often not "do I know how to use it" but "do I dare to give it access." This page breaks down WorkBuddy's security and privacy mechanisms: which data stays on your computer, which goes to the cloud, whether it can be turned off, and how to account for it to your company. The conclusion first: yes, you can entrust it with files, provided you configure the toggles and boundaries on this page properly.

What You'll Be Able to Do After Learning

• Describe WorkBuddy's data flow: file operations run locally, and snippets processed by the model are discarded after use
• Turn off the "Experience Improvement Program" to ensure data is not used for model training
• Configure permission modes and authorized directories to the state where "sensitive operations must ask you first"
• Give your company IT a checklist of materials that can pass review (compliance stance + VPC dedicated edition)

Your Current Dilemma

• You want it to organize contracts and merge reports, but fear the files being sent to who-knows-which server
• Your company's information security policy is strict: for AI tools, it's always "pass the security review first, then talk about efficiency"
• Opinions online are polarized: some say "fully local, never connects to the internet," others say "it just uploads everything to the cloud"—you don't know who to believe

When to Use This Approach

When you need to:

• Hand files containing sensitive information to WorkBuddy for processing, and be able to explain the data's destination to your company or clients
• Introduce WorkBuddy to your team and pass the information security review

And you don't want to:

• Make the call based solely on "it's made by Tencent, it should be fine"
• Discover halfway through that your data has been used for model training

🎒 Preparation Before Starting

• Have read the prerequisite course "Advanced Handbook: Assistants, Connectors, Skills, and Security" (/workbuddy/5-advanced/00-hub)
• WorkBuddy client is logged in and can send tasks normally (the facts in this article were verified against official documentation collected on 2026-11-03; applies to v5.x)
• Prepare a test folder with a few unimportant files named to "look like sensitive files" (e.g., 合同-测试.txt), to be used later to verify interception behavior

Core Idea

The official FAQ summarizes WorkBuddy's security system as "three layers of defense in depth":

Defense Layer	What It Covers	Key Mechanisms
Technical Defense Layer	Sandbox and Permissions	Bash command sandbox, file system isolation (can only access authorized directories), confirmation required for high-risk operations under default permissions, layered allow/ask/deny permissions
Management and Control Layer	Accounts and Quotas	Unified account authentication, third-party platform authorization, credit and quota control
Compliance Audit Layer	Data and Compliance	Skill security audits, skill-scanner local checks, local task execution, privacy policy

For the question "can company files be entrusted to it," what really matters is data flow. Here is the official FAQ's "data security statement" organized into a table:

Your Data	Where It Goes	Is It Stored
File operations (organizing, renaming, format conversion, etc.)	Executed in the local sandbox; original files do not go to the cloud	Stays on your computer
Conversation content and necessary snippets (in-app default model)	Uploaded to the cloud for processing	Snippets are discarded after use, not stored
Conversation content (custom model)	Passed through only to the model provider you configured	WorkBuddy server does not retain it
Whether data is used for training	Determined by the "Experience Improvement Program" toggle	Not used for training once turned off

⚠️ Don't Trust Either Extreme

"Never goes to the cloud" and "everything goes to the cloud" both contradict the official documentation. The officials state that tasks are executed locally and server-side snippets are discarded after use, while also explicitly stating: when using the in-app default model, "input and output content will be uploaded to the cloud for processing." What goes to the cloud is the content snippets needed for model processing, not your original files. Only if you can accept this boundary should you move on to the next step.

Follow Along

Step 1: Understand WorkBuddy's security data flow (what stays local, what goes to the cloud)

Why: Without understanding the data flow, every toggle you configure later will be muddled, and you won't be able to explain it to your company.

How to do it: Map your own tasks accordingly. For example, have it do batch file organization (see /workbuddy/4-scenarios/d3-file-organize)—moving, renaming, and backing up these file operations are completed locally; the content snippets it reads to understand the task go through model-side processing.

You should see: After the task completes, the files appear at the target location on your local disk, not in some cloud drive—file operations are processed locally throughout, which is the intuitive meaning of "local task execution."

• Quick check: You can explain in your own words "which part of the data goes to the cloud and which stays local"

Step 2: Confirm the permission mode so sensitive operations go through your hands first

Why: The permission mode determines whether it asks you first before modifying files or running commands — this is the setting among WorkBuddy permissions that most directly affects your security experience.

How: There are two permission modes: "Default Permissions" (sensitive operations require confirmation) and "Full Access". The entry point is the persistent permission status button at the bottom of the input box (showing "Default Permissions" by default); click it and toggle the "Allow Full Access" switch. Keep Default Permissions for daily use; when you actually switch, a "Confirm allowing full access?" dialog pops up, clearly listing the capabilities being unlocked (file operations, terminal commands, internet access) — click cancel if unsure.

You should see: Under Default Permissions, writing to sensitive paths, deleting files, running scripts, and accessing the network trigger a confirmation first, with a title like "Deleting a file hit an ask rule and requires confirmation."; clicking "View interception rules" in the top-right corner of the dialog also takes you to Settings > Security Center to see the interception reason.

💡 Don't confuse it with execution modes

Execution modes (default Agent / Plan / Ask-only) govern "how it works," while the permission mode governs "whether it can act without your consent." See the three execution modes at /workbuddy/2-daily/04-modes. Also, once Full Access is enabled it applies uniformly to all tasks in the current client and does not carry across devices; the official recommendation is to turn it off when done.

• Quick check: the current permission mode is "Default Permissions"

Step 3: Turn off the "Experience Improvement Program" to close off the path where data is used for training

Why: The official FAQ states plainly that this switch controls whether data is used for model training — this is a privacy channel you can personally shut off.

How: Open Settings > General, find "Experience Improvement Program," and turn it off.

You should see: The switch is in the off state. Official statement: once turned off, data is not used for model training.

• Quick check: "Experience Improvement Program" is off

Step 4: Narrow the authorized directories so sensitive folders stay out of the workspace

Why: File system isolation depends on the "authorization scope" itself being small enough — the larger the authorized directories, the more files it can touch.

How: Check the current workspace and authorized directories, and replace broad directories like "entire disk," "Downloads," or "Desktop" with specific project folders; do not add folders containing contracts, payroll sheets, or keys to the authorization list. Default Permissions use the workspace as the boundary, and out-of-boundary operations pause for confirmation; actions like writing to sensitive paths or bulk deletions are also intercepted to ask you first.

You should see: Test with the folder you prepared in the setup stage — have it read a "contract-like" file outside the authorized directories; it will not silently succeed but will trigger a confirmation. Also, "Sensitive Protection" is on by default: when a file genuinely contains credentials like API keys or passwords, a dialog asks "Allow the model to access sensitive credentials?" — "Allow encrypted access" is recommended, after which the ciphertext participates in the context flow.

• Quick check: the authorized directory list contains no "entire disk"-level paths and no genuinely sensitive folders

💡 Vet skills too

Skills are capability extensions that can take action, so don't install skill packages from unknown sources. There is an official Skill security audit run in coordination with the security team (tiered risk assessment), plus a local skill-scanner security check as a backstop before installation. See the previous lesson for installing and managing skills (/workbuddy/5-advanced/03-skills-market).

Step 5: Enterprise users verify compliance materials and evaluate the VPC Dedicated Edition

Why: What a corporate information security review wants is not slogans but materials that can be filed.

How: Request materials from official pre-sales using this checklist (the enterprise pre-sales entry is on the pricing page):

• User agreement and privacy agreement (officially stated to comply with the Cybersecurity Law, Data Security Law, and Personal Information Protection Law; the page footer gives formal links to the service agreement and privacy protection guide, which can be forwarded directly to legal)
• Certification evidence: the official marketing page displays ISO 27001, SOC 2, and MLPS Level III — note this is the marketing page's claim, and reviews require formal certification documents; the official FAQ additionally mentions "Tencent Cloud 400+ authoritative certification endorsement"
• For high-requirement scenarios, ask about the VPC Dedicated Edition: backend services, databases, and vector databases are deployed in a dedicated VPC, physically isolated from other tenants; enterprises connect directly from the office intranet via private links, and the official claim is data sovereignty with zero leakage
• A quick note on auditing: the Audit Center in the Security Center records interception/allow logs and supports export — useful when the security team needs an audit trail

You should see: The IT/security team receives a checklist they can verify item by item, not just the phrase "made by Tencent."

Checkpoint ✅

• You can describe the data flow: file operations run locally; under the default model, conversation content goes to the cloud and fragments are discarded after use and not stored
• "Experience Improvement Program" is off (Settings > General)
• The permission button at the bottom of the input box shows "Default Permissions," and you know in which scenarios "Full Access" should be enabled
• The authorized directories contain no sensitive folders such as contracts or keys
• The enterprise review materials checklist has been sent out (including the VPC Dedicated Edition inquiry)

Pitfall warnings

❌ Keeping "Full Access" on all the time for convenience

Reason: it applies uniformly to all tasks in the current client, and sensitive operations are no longer confirmed one by one — the official wording is "allowing full access does not mean more security." Correct practice: use "Default Permissions" day to day; officially, Full Access is recommended only for trusted tasks in isolated environments (Docker, virtual machines, temporary test directories), for short periods, and turned off when done. Unattended scheduled tasks may enable it per the official automation guide, but should be paired with narrowed authorized directories.

❌ Mistaking the "Default (Agent) execution mode" for the permission mode

Reason: both concepts have "default" in their names, making them very easy to confuse. Correct practice: execution mode governs "how it works," permission mode governs "whether it can act directly" — see the distinction at /workbuddy/2-daily/04-modes.

❌ Taking certification icons on the marketing page as compliance conclusions

Reason: ISO 27001/SOC 2/MLPS Level III are currently claims on the official marketing page; reviews require formal certification documents. Correct practice: have pre-sales provide supporting materials; if unsure, negotiate under the VPC Dedicated Edition plan.

❌ Believing the extreme claim of "fully local, never touches the network"

Reason: this contradicts the official FAQ — under the default model, input and output content is processed in the cloud. Correct practice: use the data flow table to distinguish "file operations" from "model fragments," and keep sensitive tasks within acceptable boundaries.

Advanced tips

• Custom models = zero server-side retention: the official FAQ states clearly that when using a custom model, content is only passed through to the custom model's side and the server retains no data; the trade-off is that data is instead processed by the third-party model provider (fees go to the third party and do not consume WorkBuddy credits). Suitable for scenarios where you still distrust the cloud side, but evaluate that third party first.
• Fine-grained control in the Security Center: in Settings > Security Center, files can have three-state "allow/ask/deny" rules set by path (with read, write, and delete configured separately), and rules can be set for commands by prefix and for the network by domain.
• Deletion and backup have backstops: with deletion protection on, deletions go to the recycle bin first, and bulk deletions of 50 items or more (the default) also require approval; automatic backup is on by default (a snapshot is saved before modifying files, with a backup space cap of 3000 MB); the three AI security protection switches are off by default — enable them only if needed.
• Communication and assistant channels: officially, all data transmission is HTTPS-encrypted across the full link (TLS 1.2/1.3); in the Security Center, "Transport Encryption" (an end-to-end encrypted channel between local and cloud) and "Secure Gateway" are on by default; assistant communications over WeCom, QQ, DingTalk, Feishu, etc. reuse the IM's built-in Websocket channel, opening no extra ports on the enterprise firewall and relying on no third-party tunneling tools such as frp/ngrok/Tailscale.
• When comparing with open-source solutions, note the source: the official FAQ has provided a comparison table — identity authentication (SSO single sign-on/organization structure sync/Role-based access control vs. self-built required), network security (HTTPS full-link encryption/enterprise-grade IP allowlist/secure communication gateway vs. self-configured required), permission control (member authorization/Credit limits/Skill security audit vs. manual configuration). Remember to label it "official claim" before using it in a defense.

Lesson summary

✅ On this page you have configured the four gates of WorkBuddy security:

• Data flow: file operations run locally; model snippets go to the cloud, are used once and discarded, and are not stored
• Training toggle: the "Experience Improvement Program" is turned off, and data is not used for model training
• Permissions and directories: default permissions plus narrowed authorized directories; sensitive operations must ask you first
• Enterprise track: compliance document checklist + inquiries about the VPC dedicated edition

Key configuration	Recommended value	Description
Permission mode	Default permissions	Sensitive operations require confirmation; for unattended scheduled tasks you can temporarily switch to full access
Experience Improvement Program	Off	Settings > General; once turned off, data is not used for model training
Authorized directories	Specific project folders	Bounded by the workspace; operations outside the boundary or on sensitive paths will request confirmation
Sensitive protection	Keep it on (default)	A confirmation prompt appears when credentials are matched; "Allow encrypted access" is recommended
Model selection	In-app default / custom	With the default model, content is processed in the cloud; with a custom model, data is only passed through and not retained on the server

Next lesson preview

With the security boundaries configured, you can trust it with company files. The next lesson covers "International workbuddy.ai: Differences from the Domestic Version and How to Get Started" (/workbuddy/5-advanced/05-intl). You will learn:

• Differences between the international and domestic versions in pricing and plans
• How to access the international version
• Which teams should use the international version, and whether the two versions can coexist