Field testReliability & safetyBy DanPublished Tested Updated
Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behavior
The report
Verdict
Name checked, block earned by behavior
On sample records, Guard verified one ClaudeBot claim, marked a second one an impostor and blocked neither. Only when a fake also asked for secret files did it add a one-day block, and a claim it could not check stayed pending.
On the tier list: A walkthrough of my own security code on sample data, so it places no agent on the tier list.
A crawler name is a claim, not proof
Cerberus Guard is the security system I built with Claude's help for danmanreal.com. Crawlers such as ClaudeBot and Googlebot announce themselves with a name in every request, but that name is just text, so anyone can send it. Guard looks for evidence behind the name, then decides separately whether any behavior deserves a block.
For the video, the saved October 27 version of Guard ran five times on made-up log records, with every network connection switched off. Lookups and provider lists were supplied as sample answers, so nothing here is real traffic, a real visitor or a real crawler.
Same name, two verdicts
Two sample visitors both claimed to be ClaudeBot. One came from an address on the sample provider list, and Guard marked it verified. The other did not, and Guard marked it an impostor. Neither was blocked: in this version, automatic blocking for a failed name is off by default. Spotting a fake and punishing it are two different decisions.
Search crawlers get a different check. Guard looks up the name behind the visitor's address, checks that it belongs to the expected provider, then looks that name up again; the address has to come back the same. With supplied answers the real verifier returned three different results: ok for a matching address, failed for a wrong one, and pending when the lookup service was unavailable.
Missing evidence stays pending
With no provider list to check against, a ClaudeBot claim stayed pending, not failed. That held even with automatic impostor blocking switched on and the sample clock moved past Guard's 30-minute grace window. An identity that can't be checked yet has not been proven fake, and Guard does not treat it as one.
Behavior earns the block
Then a sample visitor with a failed ClaudeBot claim asked for two secret files, the kind that hold passwords and private keys. Guard added one automatic block for one day, with the reason recorded as secret-file hunting. In the saved code, repeat offenses grow from 1 day to 7 and then 30; that ladder was read from the source, not run here.
The rule was written to local files in the sample folder. It shows Guard's decision, not a visitor being turned away on the live site.
Not locking out the owner
The last run sent the same failed claim and secret-file requests from an address configured as the owner's. Guard still logged it and still marked the claim failed, but held back the automatic block. That comes from explicit configuration, not from guessing that a familiar visitor must be me.
Where Gemini 4 Carbon fits
The video opens with Google's Gemini 4 Carbon, a model Google describes as strong at finding, validating and patching software vulnerabilities. In Google's announcement as read on October 31, access starts with a subset of cyber defenders and expands to trusted partners, with no public date. I had no access, so nothing on this page tests Carbon, and Google's benchmark scores are Google's claims, not results from this lab.
It is also a different job. A model like Carbon works on code; Guard checks visitors to a website. Claude helped write Guard, but Guard runs plain code and rules on every request, with no language model in the loop, so each decision can be traced to the check that caused it.
Limits of this report
- Sample records and supplied lookup answers only: no real visitors, no live site and no live blocking were tested.
- One run per case. The runs show how Guard decides, not how accurate it is across real traffic.
- Verification checks where a visitor comes from, not which model is behind it or whether its behavior is harmless.
- Gemini 4 Carbon was not tested; what the page says about it comes from Google's announcement.
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
Cerberus Guard, the security system for danmanreal.com (saved version of October 27, 2026)
- Same name, two visitors (ClaudeBot)
- verified · impostorone address was on the sample provider list, the other was not
- Automatic blocks for those two
- 0 · 0automatic impostor blocking is off by default in this version, so a failed name alone blocks nothing
- Search-crawler check, three offline DNS answers
- ok · failed · pendingmatching return address, wrong return address, resolver unavailable; the real verifier with supplied answers
- No provider list to check against
- pending · 0 blockseven with automatic impostor blocking switched on and the clock moved 31 minutes, past the 30-minute grace window
- Failed claim plus two secret-file requests
- 1 block · 1 dayreason recorded as secret-file hunter; the rule went into local sample files only
- Repeat offenses, from the saved code
- 1 · 7 · 30 daysread from the source, not run in this walkthrough; the optional impostor rule has its own, stricter schedule
- Configured owner address, same probes
- failed claim · 0 blocksthe owner is still logged; only the automatic block is held back
- Runs recorded, network switched off
- 5 · all finished cleanlyOctober 31 at about 6 PM ET; the demo script blocks every network connection before Guard loads
- Screen recordings scanned for private data
- 5 of 5 cleantext recognition on every second of each recording, then a visual check of each final frame
- Provider-list code used in these runs
- same file as the Oct 27 live fixbyte for byte the file my crawler report shows running on the server that day
- Context: Gemini 4 Carbon access, Google's announcement read Oct 31
- a subset of cyber defenders firstthen trusted partners through Google's Fairwind Program; no public date given. Not tested here
- Delivered episode and Shorts
- 11:12 · 1:22 · 1:11main video, then "A Crawler Name Is Not Proof" and "What Can You Use?"
Verdict
Name checked, block earned by behavior
Tier list A walkthrough of my own security code on sample data, so it places no agent on the tier list.
danmanreal.com/tests/cerberus-guard-crawler-claims
What was tested
Shorts
Related
- DMR_CERBERUS: meet the guardian
- Claude Duet ran my channel for a week and caught 269 fake AI crawlers
- Agent reliability & safety
- How I test AI tools
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