Fake AI crawlers are scanning the public web for secrets, according to a November 18, 2026 Tin Computer report. Researcher Ege Demir names two campaigns—Octopus and Gecko—that dress as GPTBot, ClaudeBot, PerplexityBot, and dozens of other documented agents while hunting .env files, cloud keys, LiteLLM admin paths, and coding-agent credentials.
The story matters because many teams allowlist AI crawlers by user agent for SEO or training transparency. Octopus exploits that habit: 32 rotating names, one Google Cloud host, hundreds of paths in seconds.
TL;DR indicators
| Scanner | Canary path / sign | Hosting pattern |
|---|---|---|
| Octopus | /z9x8c7v6b5-debug-trigger-<hostname> | Ephemeral Google Cloud VMs, 97 machines across 15 regions in the study |
| Gecko | Truncated UA ending at AppleWebKit/537.36 | Belgium Google Cloud plus long-lived DMZHOST Windows VPS |
| Tortoise (smaller) | /nonexistent-debug-trigger-7x7 | Slow scans from AS218785 |
Tin tracked them on ten sites they operate, from late September through November 18, 2026, and ties Octopus activity to prior GreyNoise, Fastly, and Cequence reports on AI-bot cosplay.
What Octopus does
API key beside an open pocket, illustrating secret paths crawlers probe
After a canary GET, Octopus fires roughly 300 secret paths in about ten seconds, rotating a different AI crawler user agent every request. Names include documented bots plus undocumented strings such as DeepSeekBot, GrokBot, and ChatGLM-Spider.
Before the burst, the same host often loads the home page as Chrome/Edge 152–157 with full browser headers, reads robots.txt and sitemap.xml, and probes Next.js and Vite manifests. Since late October, Tin says Octopus picks different path lists per framework—static HTML, Next.js, and Vite sites receive different bundles, with Vite dev-server paths such as /@fs/var/task/.env.
Mitigation by blocking one IP fails. Tin saw five regions hit one site within 25 minutes; most VMs lived under an hour.
What Gecko wants
Vault door with a key, illustrating stored agent credentials scanners hunt
Gecko is a .NET client with a 60-character user-agent truncation that drops the (KHTML, like Gecko) segment. It started on September 30 asking for Claude Code settings, then expanded through October–November to MCP JSON, .codex/auth.json, .gemini/settings.json, and LiteLLM config.
Tin says AI-tool paths rose from 4% to 10% of Gecko's sweeps by mid-November, aligning with how fast agent ecosystems add new config filenames.
Why AI teams should care
This is not classic CVE patching. It is exposure management for the same files your IDE writes locally:
/.claude/.credentials.json/.cursor/mcp.json/.codex/auth.json
If CI copies a repo—including ignored agent files—into public/ or dist/, scanners win without exploiting your app logic. The pattern parallels GTIG Claude Code settings backdoors and ChatMate indirect injection, but here the attacker never needs prompt access—only a 200 OK on a secret path.
How to protect yourself
Tin and explainx.ai align on defense in depth:
- Probe your own site for
/.env,/.git/config,/.claude/settings.local.json,/.cursor/mcp.json,/.codex/auth.json—expect 404, not a copy of your homepage. - Block dotfiles at the edge (
/.except/.well-known/). - Block Gecko's exact truncated user agent without blocking full Chrome strings.
- Verify crawler IPs against OpenAI GPTBot, Google, and other official ranges—never trust names alone.
- Rotate keys if any sensitive path ever returned real content.
For agent runtime policy, use permission boundaries like OpenAI Codex profiles and monitor tool egress.
Agent security tooling
Teams running autonomous agents should pair network rules with behavioral guardrails. AgentBeam, the agent security platform from the explainx.ai team, helps stop agents before they execute dangerous actions—complementary to blocking scanners at the edge when an agent already holds live credentials.
nginx and Cloudflare rules Tin publishes
Tin ships copy-paste examples:
- Cloudflare: block dotfiles, block Gecko's exact UA, optional
debug-triggercanary match. - nginx:
location ~ /\.(?!well-known/) { return 404; }plus UA match for Gecko.
Do not block all AI user agents to stop Octopus; you will harm legitimate crawlers and Octopus will rotate names anyway.
Connection to Black October
Tin notes Gecko's busiest day on November 17, near major vulnerability news, but does not claim causation. The structural point stands: disclosure-speed attacks and credential scanners now overlap in the same security sprint as NGINX2Root.
What developers are saying
The Hacker News thread on Tin's post debates robots.txt culture, AI-bot allowlists, and whether hosts should default-deny dotfiles. Treat commenter reports as anecdotes, not verified IOCs.
Tortoise and the shared hosting noise floor
Tin also documents Tortoise, a slower scanner using /nonexistent-debug-trigger-7x7 and HTTP/1.1 pacing. It shares budget VPS AS218785 address space with early Gecko infrastructure but appears operationally separate.
For defenders, Tortoise is a reminder that not every scan burst is Octopus. Tune detections on canary paths and framework-specific manifests, not only on AI user-agent strings.
LiteLLM and agent gateway paths on Octopus lists
Octopus requests include LiteLLM admin and spend endpoints such as /litellm/key/info and /litellm/spend/logs, added in stages through October and November. Teams running LLM proxies should assume those admin routes are on public hit lists the day they ship.
Keep gateways on private networks, enforce mutual TLS, and never expose proxy dashboards to the internet without strong auth. This mirrors lessons from Composio firewall posts: gateways concentrate keys.
Framework-specific path packs
Tin's Next.js and Vite splits are important for frontend-heavy orgs. Octopus stopped sending PHP exploit POSTs to Next sites after it detected framework fingerprints, replacing them with manifest and Kubernetes config paths.
That behavior means static export sites are not safer by default if you still publish source maps, service workers, or accidental dot directories.
Incident response checklist
If logs show Octopus or Gecko activity:
- Confirm whether any probed path returned 200 with secret content.
- Rotate API keys for OpenAI, Anthropic, Hugging Face, and cloud providers if exposure is possible.
- Purge build artifacts that copied agent dirs into
public/. - Add edge rules for dotfiles and Gecko's truncated UA.
- File abuse reports to Google Cloud with timestamps and source IPs if VMs are still live.
- Review agent sessions for anomalous tool calls after key rotation.
Why user-agent blocklists fail
ai.robots.txt and community lists now include 29 of Octopus's 32 names. Attackers benefit anyway because many site owners never verify IP ranges.
Operational policy should be: allow training crawlers only after IP verification, default-deny dotfiles for everyone else, and alert on canary paths regardless of UA.
Supply chain timing
Tin notes Gecko added paths hours before major package-compromise news but does not claim causation. The durable lesson is faster disclosure cycles and faster credential scanners now overlap—your patch Tuesday and your exposure audit cannot be separate calendars.
Mapping Tin's IOCs to your SIEM
Practical detection fields:
- URI contains
debug-triggerfor Octopus and Tortoise canaries - User-Agent equals exactly Gecko's 60-character Chrome prefix
- Request rate above 50 secret paths per minute from one IP
- AS396982 Google Cloud egress combined with GPTBot UA and non-OpenAI IP
False positives happen on security research and misconfigured monitors. Tune thresholds per property.
Relationship to ai.robots.txt community lists
Community lists merged Hunyuan and MoonshotBot names while pull-request discussion noted probe-like traffic. Octopus later impersonated those same strings.
Lesson: community blocklists are not threat intelligence. They are convenience files. Pair lists with IP verification and path behavior.
Developer laptop leakage paths
Gecko's growth into .codex/auth.json and .gemini/settings.json targets developers who sync dotfolders into web roots via sloppy deploy scripts.
Standard CI should fail builds when:
find dist public -name '.env*'returns hits- agent credential paths appear in tarballs
- source maps reference home directory paths
Align with external sandboxes for coding agents so keys never live beside static assets.
Red team exercise you can run this afternoon
Pick a staging domain. Deploy Tin's canary paths as empty 404 routes and alert if anything hits them. Run your own dotfile audit with curl from an external VPS. If you see 32 different AI UAs from one IP in ten seconds, classify as Octopus-style impersonation regardless of which name appears first.
Share results with whoever owns robots.txt and CDN WAF rules—this is not only AppSec.
Why GreyNoise and Cequence matter here
Tin cites prior GreyNoise, Fastly, and Cequence research on AI-bot cosplay. Your enterprise may already subscribe to those feeds. Add Tin canary URIs as custom signatures rather than waiting for a CVE.
Security operations centers should tag these scans as credential harvesting, not as SEO noise, so incidents escalate before keys rotate.
Tabletop with your AI vendor relations team
When logs show ClaudeBot or GPTBot UAs from unknown IPs, forward IOCs to vendor abuse inboxes and block locally. Vendors publish crawler ranges partly to help you make that distinction quickly. If abuse teams confirm impersonation, update internal threat bulletins so marketing does not approve "allow all AI bots" policies without security review.
Document false positive handling too: legitimate security scanners and uptime bots sometimes probe .env paths. Use canary plus rate together before auto-blocking entire ASNs. Keep playbooks versioned.
Related reading
- NGINX2Root HTTP/2 RCE
- Indirect prompt injection guide
- MCP tool misbinding
- 1Password for agents
- Microsoft agent framework PyPI stealer
Sources: Tin Computer, Octopus and Gecko report.
