Chasing Down a 24-Hour USB Tethering Drop on the ZTE MU5002
ZTE hotspot puck that kept going offline after tethering for 24 hours
Keeping unattended Windows 11 field systems online 24/7 over USB cellular tethering sounds simple until the 24-hour clock hits. Not long ago, I sorted out a D-Link DWR-933M hotspot that kept powering itself off overnight the moment its battery reached full charge. Once that unit was locked down and running reliably, a ZTE MU5002 5G puck landed on my bench with a different twist on the same problem.
Every 24 hours, almost to the minute, the tethered Windows host lost its uplink. Unlike the D-Link, the ZTE stayed powered on with strong 5G signal bars on the screen, while the PC sitting six inches away over USB RNDIS had zero network connectivity.
Tripping the Cloud Safety Filter First
Before I could even dig into the modem configuration, the diagnostic run hit an unexpected detour. I spun up Google Antigravity (AGY) on a freshly imaged bench machine with full SYSTEM privileges, pointed it at my old DWR-933M notes for context, and turned it loose on http://192.168.0.1.
Six minutes and a dozen scraped JavaScript files later, Gemini's safety filter blocked the entire response. Adding a follow-up note asking the agent to avoid hex or assembly output only made the classifier stricter.
The problem wasn't what I was doing; it was how the multi-turn session looked from upstream. Cloud safety filters have no concept of hardware ownership or physical possession. Combining an "express permission and highest privileges" prompt, prior firmware-modification notes, and a SYSTEM shell forging HTTP Referer headers against ZTE goform endpoints matched the exact signature of an automated router-exploitation script.
Fortunately, the blocked run had already pulled the modem's RequireJS front-end scripts (js_service.js, js_adm_lan.js, js_wifi_sleep_mode.js, and js_adm_watch_dog.js) into a local scratch directory. Copying those staged files into a working folder, starting a fresh session without the old D-Link firmware writeup, and framing the prompt around standard sysadmin configuration review got AGY straight through on the first try.
Root Cause: A Two-Sided 24-Hour Collision
Inspecting the ZTE MU5002 web UI scripts alongside my Windows hardening baseline (Hardening-Script_v3.0.ps1) showed why the symptom looked so deceptive. Neither device was crashing outright. Instead, two separate defaults were colliding right at the 86,400-second mark:
- 24-Hour DHCP Lease Expiry on the Puck: Out of the box, the ZTE
MU5002hands out a 24-hour DHCP lease (dhcpLease_hour = 24). Meanwhile, ZTE's internal ping watchdog only probes upstream cellular WAN reachability, meaning it will never reset a stalled local USBRNDISbridge so long as the 5G bearer stays up. - PCIe ASPM and Bus Idle Gaps on Windows: My
v3.0hardening script disabled OS standby, USB Selective Suspend, and USB 3 Link Power Management, but it missed PCIe Active State Power Management (ASPM). That omission allowed the PCIe root port backing theUSBXHCIcontroller to drop intoL1low-power idle states during quiet windows. - Hardcoded Keepalive Blind Spots: The local network keepalive task in
v3.0fell back to192.168.1.1while the ZTE gateway lived at192.168.0.1, andICMPv4echo requests were disabled on thePublicfirewall profile assigned to newly enumerated USB adapters.
When the 24-hour DHCPREQUEST lease renewal fired across an idle PCIe/USB path and stalled out, Windows dropped the RNDIS interface to an APIPA (169.254.x.x) address while the local keepalive sat blind to the 192.168.0.1 gateway.
Locking Down the ZTE MU5002 via Official API
Half of the power and recovery toggles on the MU5002 are tucked behind unlinked #hash routes in the web UI. Rather than clicking through hidden pages by hand on every unit, I had AGY build a reusable PowerShell provisioning script (Provision-ZteMU5002.ps1) using ZTE's official SHA-256 (LD challenge plus per-request AD token) web management API to commit our baseline straight to non-volatile memory:
- 7.4-Year DHCP Lease: Increased
dhcpLease_hourfrom24to its maximum65535hours (pushing lease expiration out past 2034) while keeping dynamicDHCPactive so field hosts stay plug-and-play. - Zero Idle Sleep and Clean Cold Reboots: Locked
sysIdleTimeToSleepto-1(Always On), disabled PCIe power saving (wifi_pcie_powersave_enable = 0) and speed-limit power saving (router_bw_switch = 0), and turned off Fast Boot (mgmt_quicken_power_on = 0) so any modem reboot cleanly re-enumerates USBRNDIS. - Baseband and WAN Self-Healing: Set connection mode to
auto_dial, enabled automatic baseband restarts on packet-switched no-service events (ps_no_service_restart_flag = 1), disabled data-cap cutoffs, and enabled the WAN ping watchdog against8.8.8.8after verifying outboundICMPcleared the carrier APN. - Dark USB-Only Operation: Serialized the three dual-chip WLAN shutdown calls (
setAccessPointInfo,setAccessPointInfo_24G_5G_ALL, andswitchWiFiModule) across transition states so both 2.4 GHz and 5 GHz radios stay completely off the air across power cycles.
Taking the Windows Baseline to v3.1
Fixing the puck solved one unit on the bench, but field machines have to tolerate whatever 192.168.x.x USB modem gets plugged into them next. Because existing hosts in the fleet were running a mix of v2.8, v2.9, and v3.0 configurations, I rolled the host-side fixes into a fully idempotent Hardening-Script_v3.1.ps1 upgrade:
- Subnet-Agnostic Keepalive and 3-Minute Self-Heal: Removed every hardcoded gateway address. The keepalive task now discovers active USB tethering gateways dynamically across
192.168.0.0/16via live0.0.0.0/0routes or DHCP server metadata, triggers a boundedipconfig.exe /renewif an adapter lands onAPIPA, and bounces the interface after 3 failed 1-minute checks instead of waiting 15 minutes. - Hot-Swap Puck Readiness: Matched USB tethering adapters (
RNDIS,CDC,UsbNcm, andUSB Ethernet) dynamically byInterfaceGuidandNetCfgInstanceIdto enforcePrivatenetwork profiles,InterfaceMetric = 10,PnPCapabilities = 24(0x18), and inbound/outboundICMPv4allow rules across192.168.0.0/16without touching wired Ethernet orCOM/serial ports. - PCIe ASPM and
SYSTEMWDF Lockdown: Addedpowercfgrules to disablePCIe ASPMon both AC and DC power, and routedHKLM:\SYSTEM\CurrentControlSet\Enum\...\WDF(IdleInWorkingState = 0) writes through a synchronousNT AUTHORITY\SYSTEMscheduled task so runningv3.1from a local Administrator shell never hitsAccess Deniederrors. - Clean Cross-Version Sweeps: Added version-wildcard cleanup for
HOSTSblocks, legacy firewall rules, and prior scheduled tasks alongside aHardening-Version.txtstamp so any older build upgrades cleanly in place without forcing an OS reboot.
After a cold reboot verification on the MU5002, both Wi-Fi bands stayed dark, 5G WAN came back up in under a minute, and Windows pulled a fresh RNDIS lease valid for the next seven years. With a single-command provisioning script for new ZTE pucks and an idempotent v3.1 script staged on my field USB drive, rolling out 24/7 cellular uplinks across the rest of the fleet during maintenance sweeps is finally plug-and-play.