Tested 23 browser agents' fingerprints: only Manus signs requests, ChatGPT no longer does
Latter-Mix2317 · reddit · 2026-10-28
The author sent 23 browser agents — ChatGPT's agent, Claude in Chrome, Cowork, computer use, Comet, Grok, Manus, Browser Use, Playwright MCP, DevTools MCP, Vercel's agent-browser and more — to the same test site with the same task, logging what the server actually sees.
Key findings:
- Manus is the only honest one: it sets Manus-User in the UA and attaches a Web Bot Auth signature to every request. ChatGPT's agent signed requests in 2025, but sent no signature in these runs.
- navigator.webdriver is false for most agents: Playwright MCP disables it, while plain Playwright scripts, agent-browser and DevTools MCP leave it on.
- Comet is inconsistent: the same public page sometimes runs in your local browser, sometimes in a cloud browser on Windows/AWS — despite claiming it only uses local for logins.
- Grok's fingerprint contradicts itself: traffic exits from the author's Mac IP and timezone, yet reports Linux x86, 8 cores, no GPU.
- Click timing can't separate agents from humans (Browser Use and Comet hold 80–450 ms, overlapping with people); what gave every clicking agent away is the pointer jumping to the target in one step instead of many small moves.
- The classic "console.log an error object and check if CDP reads it" trick detected nothing on Chrome 136–156, Playwright included.
Full user-agent list is in the comments; the author will test additional agents on request.
Related event: Only Manus Signs Honestly Among 23 Browser Agents Tested(2 posts)→
More from coding & agent
- Row-Bot: a local 27B model runs a shop's morning ops entirely on one PC via MCP — Acceptable-Object390 · 2026-10-28
- Offboarding AI agents: revoke sessions first, deleting the container isn't enough — ComprehensiveShake76 · 2026-10-28
- The compounding error problem: 95% per-step accuracy means 36% over 20 steps — KelvenSevero0 · 2026-10-28
- Dev shares AI game project's review list: laser bugs, podracer handling fixes queued — jasonkneen · 2026-10-28
- Dev builds online multiplayer castle-storming game with GPT 6.1 Terra, playable on mobile — chongdashu · 2026-10-28
- Self-distillation explained: how a model teaches itself with richer token-level training signals — arpit_bhayani · 2026-10-28