Test method: identity stability and VPS
Does the agent's cloud computer look like the same person to websites over time? Egress IP, logins and bring-your-own egress. Vendors rarely publish this, so we test it, along with whether it has a VPS (cloud computer).
Data updated: 2026-11-05 09:38Test steps
Every test is dated; products with different modes (desktop / mobile) are recorded separately.
| Step | How | What it tells you |
|---|---|---|
| 1. Check the egress IP | Have the agent open an IP lookup page (e.g. ipinfo.io) on its cloud computer, 2–3 times in one session | IP, IP type, ASN, location; whether it changes within a session |
| 2. Compare with your own device | Look up your own PC and phone egress IP at the same time | Rules out reading your own device or proxy IP by mistake |
| 3. Switch devices | Give the task from the phone app, then from the desktop app | Whether egress follows the device you use |
| 4. Toggle your own egress | If there is a “route through this computer” switch, test with it on and off, then shut the PC down while it is on | Whether you can bring your own egress; whether it goes offline or silently falls back when your device is off |
| 5. Restart the instance | Restart or reset the cloud computer, check the IP and reopen a site you were logged in to | Whether the IP changes; whether browser logins survive |
| 6. Retest the next day | Check again a day or more later | Whether ASN and country/city stay the same |
How the verdict is made
Each metric uses Tested > Official > Third-party; when sources disagree the more conservative one counts. Dynamic egress IP with no bring-your-own proxy, or logins that are not kept → public web tasks only. Logins persist and the egress IP is stable (or you can bring your own proxy / fixed egress) → OK for ordinary logins. On top of that, a residential/ISP egress or your own proxy, with no vendor warning → OK for long-term accounts. If your own egress silently falls back to the cloud egress when that device is offline, the cap is ordinary logins. Datacenter IPs, dynamic egress, offline fallback or a vendor's own warning add "not for sensitive accounts". No verdict when key data is missing.
- OK for long-term accounts
- OK for ordinary logins
- Public web tasks only
- Not for sensitive accounts added as a separate warning
Current results
| Product | Verdict | Egress IP |
|---|---|---|
| Grok Bot | OK for ordinary logins Not for sensitive accounts | Tested: Datacenter·dynamic · tested ≠ official |
| Cue(Manus) | Public web tasks only Not for sensitive accounts | Tested: Datacenter·dynamic |
| LightVela (Tencent Lighthouse) | Public web tasks only Not for sensitive accounts | Tested: Datacenter·static |
| Meta Muse | Public web tasks only Not for sensitive accounts | Tested: Datacenter·dynamic |
| Manus 2.0 | OK for ordinary logins Not for sensitive accounts | Official: Datacenter |
| Devin | OK for ordinary logins Not for sensitive accounts | Official: Datacenter·shared static |
| Replit Agent 4 | Public web tasks only Not for sensitive accounts | Third-party: Datacenter·dynamic |
Does it have a VPS (cloud computer)
- VM (sandbox): A temporary execution environment provided by the platform; its lifecycle and state may not be stable. Everyday chat and simple tasks run in a VM; every agent, chat assistants included, has one, so it is not listed.
- VPS (cloud computer): A long-lived environment whose state persists, which the user or agent can enter again and again. A VPS always includes a VM, but not the other way round; we only track whether a product has a VPS.
Tested > Official > Third-party; when sources disagree the conservative one counts (No).
| Step | How | What it tells you |
|---|---|---|
| 1. Leave a marker | On first use, have the agent write a marker file, record hostname, machine-id and uptime, and install a small tool | Gives the machine a fingerprint |
| 2. Close and reopen | Close the app or end the session, then reopen after an hour and again the next day | Marker, machine-id, installed tool and logins still there = has a VPS; gone = only a temporary VM |
| 3. Check egress IP too | Check the egress IP every time you reopen | If the machine is kept but the IP changes, that goes under identity stability |
| Product | VPS |
|---|---|
| Grok Bot | Yes |
| Cue(Manus) | Yes |
| LightVela (Tencent Lighthouse) | Yes |
| Meta Muse | No |
| Manus 2.0 | Yespaid Cloud Computer |
| Devin | No |
| Vellum Cloud | Yes |
| Zo Computer | Yes |
Terms and source labels
- Datacenter IP
- An address from a cloud data center; many sites treat it as a bot
- Residential / ISP IP
- Home broadband or carrier addresses, usually treated as a real person
- Static / dynamic
- The egress IP stays the same / changes over time or per request
- Shared static ranges
- A fixed set of addresses shared by all customers, so tests may still see the IP move within it
- Conditions in brackets
- Only holds under that condition, e.g. “desktop only”, “lost on restart”
- Source labels
- Tested = measured by TANDIU in real use (dated); Official = vendor docs; Third-party = media or provider material; No data = none of these
Vendors change egress quietly, so we retest and update results with the test date; when our test and the vendor disagree both are shown. This measures whether your own accounts get misjudged by websites when an agent uses them; it is not a guide to evading risk controls.