SASE (Secure Access Service Edge) released product feature updates and client version updates in July 2026. This topic lists those updates and provides implementation guidance for the Agent Office Security capabilities that they include.
Product feature updates
The following table lists the SASE features released in July 2026.
Module | Description | Release date | References |
Agent Office Security | Agent runtime protection supports malicious skill isolation, prompt injection attack detection, and sensitive information detection | 2026-07-30 | |
Agent permission management supports high-risk command management and data access permission management | 2026-07-30 | ||
Agent identification adds Cursor Agent, Qoder, Qoder CN, Codex, Wukong, Marvis, and WPS Comate | 2026-07-24 | ||
Data Protection | The file transfer-out policy supports priority | 2026-07-24 | |
Identity Authentication | Supports offboarding management, automatically revokes the permissions of departing employees, and supports custom offboarding cleanup policies | 2026-07-24 | |
Employee Insights adds data loss prevention (DLP) associated policies | 2026-07-24 | ||
Approval workflow | Supports integration with Feishu approval workflows | 2026-07-30 |
For guidance on rolling out the Agent Office Security capabilities in the preceding table, see Best practices for Agent Office Security.
Client version updates
The following table lists the SASE client versions released in July 2026.
Operating system | Client version | Description | Release date |
Windows/macOS | 5.2.3 | Added AI Agent runtime protection. Fixed known issues. | 2026-07-10 |
Windows/macOS | 5.2.7 | Upgraded Agent runtime protection. Added Agent control with support for Agent data permission management. Fixed known issues. | 2026-07-30 |
Best practices for Agent Office Security
Agent Office Security governs the AI Agents that employees run on managed devices. The following sections describe the risks it addresses, the governance principles behind it, the phased rollout, and the recommendations that apply throughout the rollout.
Background information
Enterprises are adopting AI Agents as a new way to get work done, which introduces security risks:
Uncontrolled shadow AI assets.
Prompt injection and prompt leakage.
Implanted malicious Skills.
Agents that access sensitive data beyond their permissions.
Compared with traditional office security risks, Agent risks have three characteristics: a high degree of automation, fast execution, and no self-correction.
To address these risks, the Agent Office Security module of SASE provides protection across asset inventory, admission control, runtime control, and audit tracing.
Management principles
Apply layered, full-lifecycle governance to Agents along one main thread: visible, controllable, and auditable.
Visible (before an incident) — Take inventory first, then control. Use AI asset scanning to identify all four asset types in your enterprise — AI Agents, LLMs, Tools, and Skills — to eliminate shadow AI and establish an asset baseline.
Controllable (during an incident) — Apply layered and flexible control. Set policies across four dimensions: people (user groups), devices (tags), assets (the Agent list), and data (classification and grading). Handling actions follow the flexible principle of observe first, then block, which balances security convergence with employee productivity.
Auditable (after an incident) — Keep a full behavior trail. Agent sessions and policy hit logs go to Log Audit, can be exported for compliance audit and tracing, and drive continuous policy iteration.
System coordination — Agent Office Security is not an isolated module. Coordinate it with Zero Trust, which governs Agent access to intranet applications and cloud AI services, and with data protection DLP, which governs the outbound transfer of the sensitive data that Agents touch. Together they form an integrated office data protection system that covers governance before an incident, audit and control during an incident, and tracing afterward.
Implementation guide
Roll out Agent Office Security in four phases. Complete each phase before you start the next one.
Phase 1: Take inventory of AI assets
Log on to the SASE console, choose Agent Office Security > AI Asset Management, and then turn on asset scanning.
In the quick scan configuration, enable Agent identification scanning, Agent security scanning (scans the Skills associated with an Agent), and session scanning (records Agent behavior logs) at the same time.
Use visual analytics to view the top five AI assets by installation count and the pending risks.
In the list, verify each category on the AI Agents, LLMs, Tools, and Skills tabs, and export the asset data to retain a baseline.
Phase 2: Enable admission control to converge entry points
Based on the scan results, define an enterprise Agent admission list: allowed, requires declaration, or prohibited.
In Agent Office Security > Agent Control Policy > Agent Disable Management, create disable policies for Agents that are not admitted or that carry high risk. Set the policy action to block startup or notify only, and configure pop-up notifications in Chinese and English to explain your corporate rules.
Use the scope of effect (user group, device tag, or device), exception users, and priority to apply differentiated admission rules.
Phase 3: Apply runtime protection to control Agent behavior
In Agent Office Security > AI Asset Management > Runtime Protection, enable the following three protections and specify their scope of effect:
Skills Protection: detects and isolates the malicious Skills that an Agent calls, which prevents plug-in and tool chain poisoning.
Prompt Protection: blocks prompt injection, jailbreak attempts, and malicious instructions.
Sensitive Data Detection: detects the sensitive data that an Agent reads in real time and links the detection to data identification rules.
In Agent Office Security > Agent Control Policy > Command Execution Control, restrict the commands that an Agent can run. You can enable built-in commands with one click or define custom commands, and set the handling action to observe or block.
In Agent Office Security > Agent Control Policy > Data Access Permissions, restrict the directories and files that an Agent can access. This protection supports blacklist and whitelist modes, custom directories (including excluded directories), and data identification rules.
Phase 4: Audit and operate for continuous iteration
In Security Operations > Log Audit, view Agent behavior and policy hit logs, and review and tune them regularly.
Coordinate with Zero Trust to manage Agent access to AI services and intranet applications. Coordinate with DLP classification and grading policies to cover the outbound transfer and tracing of Agent output.
Implementation recommendations
The following recommendations apply throughout the four phases of an Agent Office Security rollout:
Achieve visibility before control — Do not disable Agents across the board before the asset inventory is complete, because a blanket ban both misses assets and hurts employee productivity. Base the admission list on scan data.
Roll out in phases — Roll out every policy along two phased paths: pilot with selected user groups, then apply to all employees; and observe, then block. Use observation period data to calibrate false positives before you tighten the policy.
Keep exceptions to a minimum — Grant exception user status only to employees with a genuine business need, and review and revoke it regularly (quarterly as a best practice). Leave gaps between priority numbers so that you can insert policies later.
Give equal weight to control and education — Pop-up notifications support declaration. State the reason for the control and the declaration or appeal channel so that every block becomes a security awareness opportunity.
Protect data in tiers — Connect Agent data access permissions with data classification and grading, apply differentiated handling across the L1 to L4 levels, and keep confidential data invisible to Agents by default.
Close the operations loop — Review pending risks weekly, review log audit records and policy hits monthly, and update the Agent admission list and custom command rules accordingly, so that the baseline keeps pace with newly discovered Agents and Skills.