Field testTested AIBy DanPublished Tested Updated
Claude Duet ran my channel for a week and caught 269 fake AI crawlers
The report
Verdict
Found it, fixed it, waited for my yes
In one week of real work, Duet pitched this episode, flagged fake AI crawlers on my site and produced a tested fix: 34 new tests, 583 passing. It stopped before the live site; after my yes it pushed, and the fix was live a minute later.
On the tier list: A one-week field test of one product on my own work, not a coding comparison, so it places no agent on the tier list.
| Crawler name claimed | Visits | On the owner's list | Not on it | Of those, secret-file probes |
|---|---|---|---|---|
| ClaudeBot | 116 | 70 | 46 | 33 |
| Claude-User | 32 | 0 | 32 | 22 |
| Claude-SearchBot | 45 | 0 | 45 | 29 |
| GPTBot | 47 | 9 | 38 | 22 |
| OAI-SearchBot | 556 | 520 | 36 | 22 |
| ChatGPT-User | 61 | 37 | 24 | 17 |
| PerplexityBot | 38 | 10 | 28 | 15 |
| Perplexity-User | 20 | 0 | 20 | 14 |
| Total | 915 | 646 | 269 | 174 |
What I gave it
Claude Duet is Anthropic's proactive agent, opened as a research preview on October 21, 2026. It sends a morning brief, checks in every few hours and asks before it sends, pays, signs or deletes anything. I set it up like a small team: channels for YouTube, my website, my remote PC and my workshop, plus a daily routine that sends me researched video picks. One of those picks was Claude Duet itself, and that became this episode.
Out of the box, Duet can't see the files on your computer. I connected my own MCP server, DMR_REMOTE, to Claude as a custom connector. From Saturday night to Tuesday afternoon (October 24 to 27) that connection made 2,858 calls into my machine, mostly running commands and reading and writing files.
The fake AI crawlers
AI crawlers like ClaudeBot and GPTBot read sites for AI models and AI search. Their name is just a label sent with each request, so anyone can claim it. Anthropic, OpenAI and Perplexity publish the internet addresses their crawlers really use, which makes the claim checkable.
Duet flagged impostors in my site's security logs, and the week was re-checked against those published lists. 915 visits claimed one of eight AI crawler names; 269 came from addresses the owner doesn't list. 174 of those went looking for secret files such as passwords, config files and database backups. None got a real file. The table above counts visits, not distinct addresses.
The fix, and the one decision that was mine
The change it wrote checks every AI crawler claim against its owner's own list: eight crawlers, six lists, refreshed every six hours. A name with an address that isn't on the list no longer counts as that crawler. It came with 34 new tests covering all eight crawlers, and the full suite passed 583 tests with 14 skipped.
It committed the work at 11:19 AM and then stopped before the live site. After I said yes, it pushed at 12:22 PM, and the live security system restarted with the new check at 12:23. The file running on the server is byte for byte the committed file.
The real AI search traffic
The same week, OpenAI's genuine search crawler (OAI-SearchBot) visited 520 times from addresses on OpenAI's list, and ChatGPT fetched the site 37 times because a user asked it to. Separating those from the fakes is the point of the check: one kind brings readers, the other wears a borrowed badge.
Where Duet stops today
- It is a research preview on a waitlist for US Pro and Max plans, and it shares your plan's usage limits, background checks included.
- It checks in every few hours, so something urgent can arrive between checks.
- Every's review by Dan Shipper reports that scheduled runs can't reach custom connectors. That matches my setup: my routine only starts the work, and the work itself happens in a channel.
Limits of this report
- The connector log names the claude.ai connection, not the product. That the commit, push and calls came from Duet rests on Duet's own messages in my chats, which stay private.
- My routine, channels and the yes before deploy are my account from those private chats; they are not shown on this page.
- Counts cover the security log's seven-day window (October 20 to 27, 2026) and count visits, so one address can appear many times. Fake means not on the owner's published list when it was checked.
- One product, one week, one site. This is not a comparison with other agents.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
Claude Opus 5.5
- Visits claiming an AI crawler name
- 9158 crawler names from Anthropic, OpenAI and Perplexity, in my site's security log for one week (Oct 20 to 27)
- Fakes: address not on the owner's list
- 269every Claude-SearchBot and Claude-User visit was one; 46 of 116 ClaudeBot visits
- Fakes hunting for secret files
- 174 · 0 got a real fileevery answer was not found, forbidden, gone or a redirect
- The fix: crawlers checked against owner lists
- 8 crawlers · 6 listseach list refreshed every 6 hours; Anthropic publishes one list for its three crawlers
- Tests
- 34 new · 583 passingfull suite: 583 passed, 14 skipped
- Committed, pushed, live (Oct 27, ET)
- 11:19 · 12:22 · 12:23the file running on the server matches the committed file byte for byte
- Real OpenAI search crawler visits
- 520OAI-SearchBot from addresses on OpenAI's list; ChatGPT also fetched the site 37 times for users
- Claude connector calls into my PC
- 2,858Oct 24, 10:54 PM to Oct 27, 4:16 PM ET; 1,441 on Oct 26
- Delivered episode
- 525.033 s
Verdict
Found it, fixed it, waited for my yes
Tier list A one-week field test of one product on my own work, not a coding comparison, so it places no agent on the tier list.
danmanreal.com/tests/claude-duet-runs-my-channel
What was tested
On the tier list's bench
- Claude Opus 5.5 · 5 receipts on file
Shorts
Related
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4