Unsourced
Field data

What fake AI bots are really after: secret files, more than pages.

We traced 34,801 visits that claimed to be AI bots. Four in ten fake visits asked for secret or config files like .env and cloud keys. Not one real bot did.

Fake AI bots are not only scrapers copying pages. Of 34,801 visits that claimed an AI-bot name, 9,951 failed the identity check. Four in ten of those asked for secret files: .env, cloud keys, git folders. They did it under the names of ChatGPT, Claude and Perplexity. The 18,491 genuine bot visits asked for none of it.

Published 17 November 2026 · Unsourced Research · All research

How we decided what is fake

We took every visit to unsourced.app between 10 June and 11 November 2026 that claimed to be one of 28 named AI crawlers. That is 34,801 visits from 3,315 IP addresses. We checked each address against the IP list the operator publishes, and against forward-confirmed reverse DNS. Then we sorted the visits three ways.

Genuine: 18,491 visits (53%) passed the check. Fake: 9,951 (29%) had proof against them: the address sat outside the operator’s published ranges, its confirmed hostname belonged to someone else, or the name belongs to nothing that crawls. Unverifiable: 6,359 (18%) had no decisive check available, so we do not call them fake. Our last post counted everything unproven. This one calls a visit fake only when we can point at the proof.

1. Four in ten fake visits asked for secret files

Real bots asked for pages. Many fakes asked for things no public site should serve. Of 9,951 fake visits, 4,045 (41%) were probes: requests for secret, config or admin files. They used 648 different paths. Genuine bots made 0 such requests in 18,491 visits. Of all 5,105 probe requests, 4,045 came from fakes and 1,060 from visits we could not check.

unsourced.app What the fakes asked for Probe requests by fake AI-bot visits, by kind of file Env files2,917 Cloud keys374 Config files316 Git repos229 PHP files108 Dev servers87 WordPress14 4,045 probe requests from fake visits, 10 Jun to 11 Nov 2026. The most requested path was /.env.

The pattern points to a search for credentials. /.env and its variants (.env.local, .env.production, .env.bak) made up 2,917 of 4,045 probes. Next came cloud credentials such as /.aws/credentials, then config files, then /.git/config. Anyone who finds one of these exposed gets access to the account behind it. Such requests would reach those files if they existed. We cannot see who sent them or why, only what they asked for.

2. They wear the names of the bots you would let in

A scanner that calls itself ChatGPT-User gets waved through by site owners who want AI visibility. The name with the most fake visits was ChatGPT-User: 2,579 of its 3,674 visits (70%) came from addresses OpenAI does not list.

unsourced.app How much of each name was fake Share of visits under each bot name that failed the identity check Google-Extended100% ChatGPT-User70% CCBot66% PerplexityBot59% Bingbot56% GPTBot42% OAI-SearchBot38% Amazonbot34% ClaudeBot18% Visits to unsourced.app, 10 Jun to 11 Nov 2026. Bot names with fakes and at least 250 visits. Fake = proven, not merely unverified.

Four other names failed every single check: Claude-User (144 visits), OAI-AdsBot (142), Perplexity-User (133) and Claude-SearchBot (113). These are real bots run by their operators. None of the visits under those names came from the operator. DuckAssistBot had 110 fake visits out of 115.

3. Most of it came from rented servers on one cloud

6,954 fake visits (70%) came from 38 addresses on Google Cloud customer servers. They first showed up on 19 July. Between them they wore 16 bot names, and 3,445 of their visits (50%) were probes. That is 85% of all probe requests from fakes. Across all 651 fake addresses, 63 wore five or more names and the most any one wore was 13. Most (583 of 651) wore just one.

That fits a scanner that borrows crawler names to look harmless. We make no claim about who runs it. Google Cloud is a rented platform, and the same platform hosts plenty of honest services.

4. Two names that nothing sends

Google’s crawler list says Google-Extended has no separate user agent string. Apple’s Applebot page says Applebot-Extended does not crawl webpages. Both are robots.txt tokens for site owners to opt out with. They are not visitors. Yet we logged 474 visits named Google-Extended and 9 named Applebot-Extended. All 483 were fake by definition. If you see either name in a log, someone typed it in.

5. Not every fake was after secrets

About 6 in 10 fake visits were not counted as probes. 1,558 asked for our homepage, bot guides, research posts or studies. These look like scrapers, or monitoring tools, or just curiosity under a borrowed name. We cannot tell which from a log line. Another 1,578 asked for .json or .csv files, such as /credentials.json and /service-account.json. We left those out of the probe count, so four in ten may be an undercount. Fake bots that read ordinary pages are the harder case, because nothing in the request looks wrong. Identity is the check that catches them. We covered that in identity or path.

What came back

Nothing sensitive. On 14 November we requested seven of the paths named in this post (such as /.env, /.aws/credentials, /.git/config and /service-account.json) ourselves. Each returned our ordinary app page or a 404, never a file, because no such file is in our public folder. Status codes were logged only from 30 October, so we cannot show the same for every earlier probe. We do not publish what the probes did next. The risk for any site is a mistake: a stray backup, a build folder left in place. Block these paths at the server, so one slip cannot turn into a leak.

What to do about it

Do not trust the name. A user agent is a claim. Check the address against the operator’s list. Verify a bot does it in one step, and verify before you block covers the method.

Block secret paths for everyone. Return 404 for dot-files (except /.well-known/), .git, .env and config files at the web server. No real AI bot needs them.

Watch for one address wearing many names. It is a strong sign of a fake. Most fake addresses in our data (583 of 651) used just one name, so this will not catch them all.

Do not block AI bots by name alone. The name tells you little. Blocking GPTBot does not stop a scanner calling itself GPTBot, and allowing it does not make a scanner safe. More in blocking AI crawlers is not one switch.

Limits

These are visits to unsourced.app only, over five months. A site about AI crawlers likely draws more fakes than most. We checked identity against today’s IP lists and live DNS, rerun with retries. Operators change their ranges, so a few older visits may be judged wrongly. Counts moved a little between reruns of the DNS check, so figures can shift by a few hundred visits between runs. Bytespider is counted as unverifiable: we found no official ByteDance page that documents how to check it. 6,359 visits had no decisive check and are not counted as fake. Visits that used a real bot’s name from an unlisted address are fake only in the sense that they did not come from the operator, and we do not know their intent. Operator statements are as we read them on 14 November. Anthropic began publishing its IP ranges in 2026, so older Claude visits are judged against today’s list. Status codes were logged only from 30 October.

Common questions

What do fake AI bots want?

On our site, secret and config files were the biggest single group, but not the majority. Of 9,951 fake visits, 4,045 (41%) asked for secret or config files. Examples are .env, cloud credentials, git folders and admin paths. They used 648 different paths. Another 1,558 asked for our homepage, bot guides, research posts or studies. Genuine bots made 0 such requests in 18,491 visits. We cannot see intent, only what was asked for.

Which AI bot name had the most fake visits?

ChatGPT-User. Of 3,674 visits that used that name, 2,579 (70%) came from addresses OpenAI does not list. Several smaller names were fake on every visit, such as Claude-User and Perplexity-User. CCBot (66%) and PerplexityBot (59%) came next. ClaudeBot was faked least among the big names, at 18%. These are shares of visits to one site, not a ranking of how often each name is faked everywhere.

Are Google-Extended and Applebot-Extended real crawlers?

No. Google says Google-Extended has no separate user agent string, and Apple says Applebot-Extended does not crawl webpages. Both are robots.txt tokens. We logged 474 visits named Google-Extended and 9 named Applebot-Extended. All were fake, and 461 of them ran from Google Cloud customer servers.

How can I tell a fake AI bot from a real one?

Do not trust the name. Check the visiting IP against the list the operator publishes, and use forward-confirmed reverse DNS where the operator documents it. Real bots on our site never asked for secret files. A visit that asks for /.env under a bot name is a fake or a scanner. You can check an address with our Verify a bot tool.

Did the fake bots get anything from the site?

We found none. When we requested paths like /.env ourselves on 14 November, we got our normal app page or a 404, never a file. The server holds no such files in the public folder. But the lesson holds for any site: block these paths at the server, so a mistake in a deploy cannot expose one. Status codes were logged only from 30 October, so we cannot show the same for every earlier probe.

Related

AI crawlers exposed

What the genuine bots really do on a site, bot by bot.

Read it →

Identity or path?

Two ways to clean an AI-crawler log, run on one log.

Read it →

Verify a bot

Check any IP against an AI operator’s published ranges.

Read it →

See the evidence for your own site

Unsourced captures which AI assistants cite you, proves which crawlers really fetched your pages, and re-checks after you act: evidence, not a score.

© Unsourced, the evidence layer for AI search. Our data and findings are free to quote and cite. Please attribute to Unsourced and link to unsourced.app.

Crawler test: this link is blocked in our robots.txt.