deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (hnrss.org)

Scanners impersonating GPTBot and ClaudeBot probe sites for .env files and AI keys

A study of ten websites documents three scanners that spoof AI crawler user agents to hunt exposed .env files, cloud credentials and AI tool tokens, and lists the log signatures that identify each one.

Scanners impersonating GPTBot and ClaudeBot probe sites for .env files and AI keys

A report published on tin.computer by Ege Demir, and surfaced on the Hacker News front page, documents three scanning campaigns that masquerade as AI crawlers while hunting for exposed secrets on websites. The author monitored ten sites from late September to 18 November 2026, named the scanners Octopus, Gecko and Tortoise, and distilled each one into a searchable log signature.

Octopus impersonates 32 AI crawlers

Octopus announces itself with a request for a path starting /z9x8c7v6b5-debug-trigger- and ending with the target's own hostname. Two random paths follow, which always return 404; the author believes they calibrate how a site reports missing pages, so the scanner can distinguish a genuine .env file from a not-found page served with a 200 status. Roughly 300 requests for secret files then land within about ten seconds, each under a different user agent drawn from a pool of 32 AI crawler names.

Seven of those names are documented by no company, including DeepSeekBot, GrokBot, Qwenbot, MoonshotBot and Bravebot. The other 25 copy real crawlers such as GPTBot, ClaudeBot, PerplexityBot, Meta-ExternalAgent, Applebot, Bytespider and Googlebot, mostly letter for letter. The giveaway, according to the report, is that all 32 variants come from a single machine on one connection within seconds, behaviour no legitimate crawler shows.

The infrastructure churns constantly. Each visit came from a fresh Google Compute Engine instance (AS396982, *.bc.googleusercontent.com), almost always in a new region. The author counted 97 machines across 15 regions during the study window; 95 touched only one website, 81 appeared on a single day, and 65 were gone within an hour. On 24 October, five machines in five regions scanned the same site within 25 minutes, which makes address-level blocking futile. Octopus also works from a hostname list rather than an IP sweep, since it reached hostnames that have no certificate of their own.

It browses like a browser before it probes

About 1.5 seconds before the canary request, the same machine loads the home page with an ordinary Chrome or Edge user agent (versions 152 to 157) and full browser headers. It reads robots.txt and sitemap.xml, follows a median of 68 links, and requests build manifests such as /.next/build-manifest.json, /ngsw.json and /.vite/manifest.json, plus /admin, /settings/api-keys, /graphql and /swagger.

Since late October that reconnaissance feeds a second stage: each site receives a path list matched to its framework. PHP, Python, Laravel, Terraform and Docker targets get exploit-style POST requests; Next.js sites get build files and Kubernetes configs; Vite sites get dev-server paths such as /@fs/var/task/.env. The full catalogue runs to 457 paths, covering .env variants, AWS and gcloud credentials, service-account JSON, LiteLLM key and spending endpoints, Flowise and Langflow endpoints, HuggingFace tokens, Vercel and CI configuration, Vault paths and Spring Boot actuator routes. Around half matches public wordlists; the rest looks purpose-built. The list also evolves: Flowise and Langflow exploitation paths appeared on 20 October, two weeks after advisories for both tools, and were dropped a day later.

GreyNoise, Fastly and Cequence have previously reported scanners with fake AI crawler names on Google Cloud. Cequence's summary said the scanner ignored ordinary pages; the author's logs disagree, but those page requests carry browser user agents, so filters keyed to crawler names never see them.

Gecko and Tortoise

Gecko is a .NET program running mostly from Google Cloud's Belgium region, with some rented Windows machines at an offshore hoster. Its user agent is a Chrome 128 string cut off at 60 characters, which removes the trailing like Gecko token, leaving Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36. It asks for the files where Claude Code, Codex, Cursor and Gemini CLI store credentials, including /.claude/settings.json and /.cursor/mcp.json, and its path list grew from 4 to 26 entries in five weeks. Tortoise is smaller and slower: a Chrome/126 user agent on Windows, roughly one request every few seconds, originating from TC DATACENTER (AS218785). Who runs the campaigns is unproven, though the evidence points to two separate operators behind Octopus and Gecko.

Signatures to search for

The report reduces each scanner to greppable signals:

  • Octopus: any hit on /z9x8c7v6b5-debug-trigger- followed by your hostname, or many different AI crawler names arriving from one address within seconds.
  • Gecko: the truncated AppleWebKit/537.36 user agent paired with requests for /.claude/settings.json or /.cursor/mcp.json.
  • Tortoise: the path /nonexistent-debug-trigger-7x7.

Why it matters

These campaigns exploit the fact that AI crawler traffic has become normal background noise on the web. Teams that tolerate or allowlist GPTBot and ClaudeBot give the impersonators cover, and the browser-user-agent reconnaissance stage slips past any filter that only watches crawler names. Because each target gets a fresh, short-lived cloud VM, IP blocking is largely useless; canary paths and behavioural signatures are the practical defence. The targets themselves — LLM gateway keys, provider tokens, CLI credential files, cloud credentials — are directly monetizable, so anyone whose .env file or key store is accidentally web-accessible should assume a scanner will find it, probably within weeks.

  • #security
  • #web-scanning
  • #api-keys
  • #ai-crawlers
  • #google-cloud

Related posts