How Tencent WorkBuddy 5.0’s Security Center Makes AI Office Work Safer

WorkBuddy 5.0’s Security Center is Tencent’s attempt to make desktop AI agents safer at the point where they act on files and the operating system—not only when they interpret a prompt. In versions 5.0 and later, the center brings sandboxing, deletion recovery, and edit-version restoration into one place, with the core controls enabled by default. Open the client’s lower-left avatar, then go to Settings → Security Center. 612

The core safety model: constrain execution, not just instructions

Under Default Permission, WorkBuddy routes agent commands through a security sandbox before execution. Routine work within the workspace can proceed, while actions such as disk formatting, accessing sensitive system locations, or traversing the system root can be intercepted. 69

That matters because a filter based only on command text is limited to recognizing known strings or patterns. WorkBuddy describes its approach as making decisions at the operating-system API layer, based on the operation the agent is attempting. In practical terms, the protection is intended to remain relevant even if a risky action is expressed differently from a previously known command. 67

The sandbox is an additional control, not a substitute for endpoint protection, data classification, or an independent backup plan. Tencent’s enterprise guidance recommends combining a dedicated workspace, default permissions, backups for important files, and human review. 5

Default Permission vs. Full Access

The two modes represent different trade-offs between autonomy and oversight:

• Default Permission is the recommended everyday mode. It allows ordinary work but can ask for confirmation when a task involves protected paths, important or bulk deletion, scripts, commands, external applications, or network-related risk. The prompt is intended to let the user review the action, target, and likely scope before proceeding. 25
• Full Access turns off those step-by-step confirmations for high-risk actions, including file writes and deletion, scripts, and external-program calls. It is best reserved for a short, trusted task in an isolated environment where recovery is possible—not for production, customer, financial, or irreplaceable data. 28

For a safer workflow, keep the original important files outside the task workspace, give the agent copies to process, and return to Default Permission when a one-off elevated task is complete. 18

Recovering from mistaken deletion or edits

Security Center addresses two common failure modes of autonomous file work.

Deleted files: Agent-initiated deletion is configured to use the system recycle bin rather than immediately permanently removing the file, making recovery possible through the operating system’s normal recycle-bin workflow. 69

Edited files: Before the agent modifies an existing file, WorkBuddy can save a backup of the original. That supports restoration of a prior version of PowerPoint, Word, image, and PDF files—not merely recovery of extracted text. New files are not duplicated as pre-edit backups because there is no earlier file version to preserve. 29

The backup store has a 1 GB default capacity. Older backup versions are automatically cleared as the store fills, so the convenience feature should not be treated as long-term archival storage. 19

At the time these features were announced, automatic backup and version restoration were fully available on Windows; the equivalent Mac capability was described as forthcoming. 9

Encryption and generated-content safeguards

WorkBuddy documentation states that data transmission uses HTTPS with TLS 1.2 or TLS 1.3, intended to protect data in transit against eavesdropping, tampering, and man-in-the-middle attacks. 17

For enterprise use, Tencent also describes an AIGC content-safety engine that detects and filters generated content identified as noncompliant, harmful, or otherwise unsafe. This is an output-control layer, separate from the local execution safeguards that govern what an agent can do on a device. 2128

What this approach changes for AI-agent work

The Security Center combines three layers that solve different problems:

1. Execution boundaries limit or interrupt risky OS and file actions.
2. User approval keeps people in the loop for sensitive operations under Default Permission.
3. Recovery controls make accidental deletion and unwanted edits less final.

These controls reduce risk, but they do not make an agent infallible. The most defensible setup remains a separate workspace, least-privilege permissions, human review before consequential actions, and independent backups for anything that cannot be replaced. 58