Fake crawlers and exposed secrets
bots that pose as AI crawlers to look for secret filesScan your site for exposed secrets
We ask your site for the same secret files these bots ask for, and show which ones answer. 75 read-only checks. We never keep secret values and never publish results.
44% of requests signed GrokBot were fake, where we could check
Each square is 1% of the requests that named the crawler, checked against its owner's published addresses.
Most of what they ask for is API and admin config
304,327 requests for secret paths, by what they target. This uses a wider path list than the headline figure.
The most trusted names get faked most
Fake requests for secret files, by the name they used. 12 Mar to 29 Jun 2026.
CCBot16565GrokBot11458DeepSeekBot10455Two tools, each with its own tell
The fake headers split into two templates on two networks.
Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ClaudeBot/1.0; +mailto:[contact omitted]Gives an email contact where the real crawler gives its official page. We omit both addresses.
Mozilla/5.0 (KHTML, like Gecko; compatible; GPTBot/1.3; +https://openai.com/gptbotThe bracket after Mozilla never closes. Real OpenAI agents always close it.
One address, 16 names, 2.3 seconds
25 Sep 2026: 104 requests from one Google Cloud address, a new crawler name almost every time. None got a page back.
14:41:04.666ChatGPT-UserGET /api/config30814:41:04.780PerplexityBotGET /env.js40414:41:04.788MistralAI-UserGET /env.json40414:41:04.794Claude-UserGET /firebase-config.json40414:41:04.882cohere-aiGET /admin%2f.env40414:41:04.930GrokBotGET /dashboard%2f.env4046 of the 18 requests we publish, of 104 in the burst.
How we measure
We read server logs from 127 sites that share them with Trakkr, 1 Aug to 14 Oct 2026. A request counts as fake when it names an AI crawler but comes from outside that crawler's published addresses (for GrokBot, when the same address also claimed a rival's crawler). Only crawlers we can check this way appear in the waffles, so they show 4.
62% of the fakes hit one site; without it, 3.8% of checkable crawler traffic was fake. Logs record status codes, not what was sent back: 10,078 probes got a 200 on 20 sites, mostly catch-all pages. The name list and the two tools come from 12 Mar to 29 Jun 2026. Addresses are withheld.
Is GPTBot really crawling my site?
Maybe not. Anyone can put "GPTBot" in a user-agent header, and bots do exactly that to get past the rules sites set for AI crawlers. The only way to be sure is to check that the request came from an address OpenAI publishes, not just trust the name. Where the address could be checked, 5.7% of GPTBot requests between 1 Aug to 14 Oct 2026 did not check out.
How do I verify a real AI crawler?
Three layers. Never count a request for a secret file (like /.env) as a crawl. Trust Cloudflare’s verified-bot signal where it is present. Otherwise check the client IP against the operator’s published ranges (openai.com/gptbot.json and the equivalents for Anthropic and Perplexity) or a reverse-DNS lookup. Trakkr counts a crawl only when the source checks out.
Did any of these scans succeed?
The logs cannot say for sure, because they record status codes, not what a site sent back. Between 1 Aug to 14 Oct 2026, 10,078 secret-file probes got a 200 on 20 sites. Most of those were catch-all pages that answer 200 to any path. On 7 sites only a few secret-looking paths returned a 200, and logs alone cannot clear those. Run the scanner on this page to check your own site.
How big is the problem?
Where the client IP could be checked, 7.8% of AI crawler requests were fake between 1 Aug to 14 Oct 2026 (6.7M requests on 100 sites). One site took 62% of the fakes: without it the figure is 3.8%, and the median site saw none. Some names are hit much harder than others: 43.8% of GrokBot requests were fake, against 11.7% of PerplexityBot, 7.4% of ClaudeBot and 5.7% of GPTBot.