CyberSec

Two things live here: a running read on the automated scanning traffic this site receives, sorted by what it's actually after — and a record of security research I've submitted through formal disclosure channels.

What's probing this site

Window: 2026-09-20 → 2026-09-23 (72h) · source: Cloudflare GraphQL Analytics API · last updated 2026-09-23

Where it's coming from

Netherlands — 627 hits Taiwan — 569 hits United States — 481 hits Germany — 314 hits Belgium — 170 hits India — 93 hits Japan — 73 hits France — 72 hits Romania — 67 hits Hong Kong — 61 hits Singapore — 47 hits Canada — 45 hits United Kingdom — 44 hits Italy — 39 hits Qatar — 36 hits Brazil — 28 hits Luxembourg — 15 hits Hungary — 14 hits Finland — 12 hits Indonesia — 11 hits Australia — 9 hits South Africa — 8 hits Lithuania — 4 hits Vietnam — 3 hits Chile — 3 hits Togo — 2 hits Sweden — 2 hits Tunisia — 2 hits Morocco — 1 hits Turkey — 1 hits Switzerland — 1 hits

31 countries, 2861 hits mapped · bubble size and opacity scale with hit volume · hover a bubble for exact counts. 7 additional hits routed through Tor exit nodes aren't plotted — Tor has no fixed geographic source.

6,226total requests
1,131exploit / recon hits*
11attack categories
~40distinct source ASNs
0files that existed

Measured against a 1,983-row sample of the 6,226 total requests in the window — Cloudflare's per-request log is sampled on this plan, so category counts are representative, not an exact census.

Environment & config-file harvesting high impact 432 hits · the single largest category
Target

.env in every naming convention a framework might use — .env.production, .env.backup, .env.local, .env.save, .env.swp (an editor swap file) — plus config.json, config.yaml, and framework-specific settings files.

Mechanism

On real Node, Laravel, Django, or Rails apps, .env commonly sits in the web root by accident — copied there by a lazy deploy script, or left readable because the framework's default .gitignore doesn't stop a misconfigured web server from serving dotfiles. No exploit is needed; it's a GET request against a filename people forget isn't supposed to be public.

What success looks like

Database connection strings, third-party API keys (Stripe, AWS, SMTP), session-signing secrets, sometimes OAuth client secrets — the entire trust boundary of an application in one plaintext file.

/.env
/.env.production
/.env.backup
/.env.staging
/api/.env
/.env.local
/app/.env
/.env.old
/.env.dev
/.env.save
Version-control directory exposure high impact 52 hits · concentrated on two files
Target

/.git/config (23 hits) and /.git/HEAD (14 hits) overwhelmingly, plus scattered probes for /.git/index, /.gitmodules, and /.svn/entries.

Mechanism

If a deploy process copies a project's working directory straight into the web root — instead of a build artifact — the .git folder ships too. HEAD and config are the first two files a tool like git-dumper reads to confirm a live repo is there, before walking .git/objects to reconstruct the entire commit history file by file over plain HTTP.

What success looks like

Full source code, including anything ever committed and later “deleted” — old credentials in commit history are a classic finding, since deleting a file doesn't remove it from git's object store.

/.git/config
/.git/HEAD
/.git/index
/.gitmodules
/.svn/entries
SSH keys & cloud credential harvesting high impact 31 hits
Target

/.ssh/id_rsa, /.ssh/id_ecdsa, /.ssh/known_hosts, /.ssh/authorized_keys, /.netrc, /.npmrc, /.aws/config, /.git-credentials, /.bash_history.

Mechanism

These paths assume a container or VM image was built carelessly — a developer's home directory baked into a Docker layer, or a CI runner's SSH agent state left on disk and accidentally exposed by a static file server pointed at the wrong root.

What success looks like

A private key is a credential that grants lateral movement — access to other servers, git remotes, or cloud accounts that trust that key, not just the one machine it was found on. This is the category attackers chain into ransomware and supply-chain footholds.

/.ssh/id_rsa
/.ssh/id_ecdsa
/.ssh/known_hosts
/.netrc
/.npmrc
/.aws/config
/.git-credentials
/.bash_history
WordPress fingerprinting & plugin CVE probing med–high impact 260 hits · second-largest category, one dominant path
Target

/wp-json/gravitysmtp/v1/tests/mock-data alone accounts for 75 hits — a REST route belonging to the Gravity Forms SMTP plugin, subject of a public unauthenticated-disclosure writeup. The rest is generic WordPress fingerprinting: /wp-login.php, /xmlrpc.php, and dozens of /<guess>/wp-includes/wlwmanifest.xml requests trying subdirectory install locations.

Mechanism

Mass scanners keep a running list of every WordPress plugin with a disclosed CVE and sweep the internet for the tell-tale REST route or file. wlwmanifest.xml is a relic of the old Windows Live Writer blogging client — its presence confirms “WordPress lives here,” queuing the target for a more targeted second pass.

What success looks like

Ranges from information disclosure (the GravitySMTP route could leak test-email content) up to full remote code execution for older, unpatched plugin vulnerabilities.

/wp-json/gravitysmtp/v1/tests/mock-data
/wp-login.php
/xmlrpc.php
/blog/wp-includes/wlwmanifest.xml
/wp1/wp-includes/wlwmanifest.xml
/2019/wp-includes/wlwmanifest.xml
PHP info-disclosure & backdoor probes medium impact 52 hits
Target

/phpinfo.php, /i.php, /pi.php, /test.php, /debug.php, plus specific ones like /icecoder/lib/terminal-xhr.php (a known vulnerable web-based code editor) and /set_captcha_validated.php.

Mechanism

Two goals share this bucket: find a stray phpinfo() call, a diagnostic dump of the server environment someone forgot to delete; or check for a webshell left behind by a previous, successful compromise of the same box. Scanners routinely probe for other attackers' backdoors, not just fresh vulnerabilities.

What success looks like

A live phpinfo() page is reconnaissance gold for planning the next exploit against that exact PHP build. A live webshell is immediate code execution with someone else having already done the hard part.

/phpinfo.php
/i.php
/pi.php
/test.php
/debug.php
/icecoder/lib/terminal-xhr.php
/set_captcha_validated.php
Framework debug & monitoring endpoints high impact 34 hits
Target

Spring Boot's /actuator/env, /actuator/heapdump, /actuator/mappings; Laravel's /telescope/requests; IIS's /trace.axd and /elmah.axd; and Vite's /@fs/proc/self/environ and /@vite/env.

Mechanism

Legitimate developer tools, safe in local development and catastrophic in production. Spring's Actuator suite is designed to expose internals for operators — /actuator/heapdump literally downloads a snapshot of the JVM's live memory. Vite's @fs is its dev-server file-serving escape hatch, with a documented class of path-traversal advisories when a dev server is left reachable from the internet.

What success looks like

A heap dump can contain session tokens, database passwords held in memory, and API keys. trace.axd/elmah.axd leak recent request/response bodies and unhandled exceptions, often including stack traces with connection strings.

/actuator/env
/actuator/heapdump
/actuator/mappings
/telescope/requests
/trace.axd
/elmah.axd
/@fs/proc/self/environ
/@vite/env
Enterprise CMS & app-server fingerprinting high impact (historically) 20 hits · low volume, high pedigree
Target

/login.action (Apache Struts2's default login handler), a Jira/Confluence-flavored path ending in META-INF/maven/com.atlassian.jira/.../pom.properties, Joomla's /administrator/ and its JCE editor component, Microsoft Exchange's /ecp/Current/exporttool/..., and cPanel/WHM subdomain-proxy detection paths.

Mechanism

Each is a name-brand enterprise product with a history of severe, publicly documented vulnerabilities. /login.action is the same class of Struts2 endpoint behind the 2017 Equifax breach (CVE-2017-5638) — scanners still check for it eight years on. The Jira path is a version-fingerprinting trick: reading pom.properties discloses the exact build number.

What success looks like

Struts2/Jira/Exchange RCEs of this era are typically full remote code execution as the application server user — the ceiling outcome for a web vulnerability.

/login.action
/.../pom.properties
/administrator/
/administrator/components/com_jce/
/ecp/Current/exporttool/
Cloud & container infrastructure exposure high impact 18 hits
Target

/v2/_catalog (Docker Registry HTTP API), /.kube/config, gcp-key.json, /aws-exports.js, /payment_gateways/stripe.yaml, /.dockercfg.

Mechanism

These check for infrastructure tooling exposed to the public internet instead of an internal network — an unauthenticated Docker registry lets anyone list and pull every image ever pushed; a leaked kubeconfig hands over cluster-admin access outright.

What success looks like

Pulled container images routinely have build secrets baked into intermediate layers even after the final image looks clean — a well-known Docker antipattern.

/v2/_catalog
/.kube/config
/gcp-key.json
/aws-exports.js
/payment_gateways/stripe.yaml
/.dockercfg
Path traversal & LFI depth-fuzzing med–high impact 29 hits · one distinctive campaign
Target

A single source ran an escalating sequence: //images/images/cache.php, then //images/images/images/cache.php, up to eight nested images/ segments deep, alongside a cgi-bin/ variant and percent-encoded traversal like /%2f%2eenv.

Mechanism

Older PHP thumbnail/cache scripts (the timthumb.php-style vulnerability family) build a filesystem path by naively concatenating the URL onto a base directory. Repeating a directory segment tests how many levels of traversal a target's path sanitization tolerates. Percent- and double-percent-encoding is the standard trick for slipping ../ past a naive string-matching filter.

What success looks like

Arbitrary file read from outside the intended web root — often used to reach /etc/passwd, application config files, or /proc/self/environ, which on Linux contains every environment variable of the running process.

/images/images/cache.php
/images/images/images/images/cache.php
/cgi-bin/cgi-bin/cgi-bin/cgi-bin/cache.php
/%2f%2eenv
Generic API & GraphQL reconnaissance low impact (recon only) 50 hits
Target

/graphql, /api/graphql, /v1/graphql, /api/gql, bare /api, and /api/session/properties (27 hits alone from a small cluster of IPs).

Mechanism

Almost every modern SaaS product exposes a GraphQL or REST endpoint at one of a handful of conventional paths. This is broad-spectrum fingerprinting: find the endpoint first, introspect its schema second, look for authorization bugs third. Against a static site these all return 405 Method Not Allowed.

What success looks like

An open GraphQL introspection query is often the fastest way to map an entire application's data model and privilege structure in one request.

/graphql
/api/graphql
/v1/graphql
/api/gql
/api/session/properties
Emerging: AI-agent manifest discovery informational (new) 5 hits · small, but worth watching
Target

/.well-known/agent.json, /.well-known/agents.json, /.well-known/agent-card.json.

Mechanism

These are the discovery paths defined by emerging agent-interoperability standards (Google's Agent2Agent protocol and similar “agent card” conventions) for a site to advertise an AI agent's capabilities and endpoints. Their appearance mixed into general-purpose exploit-scan traffic suggests scanner toolkits are already being updated to fingerprint sites exposing autonomous agents.

What success looks like

An exposed agent manifest could reveal callable tool endpoints, API surfaces, or authentication schemes for an AI agent before any conventional vulnerability is even in play — a genuinely new reconnaissance target class.

/.well-known/agent.json
/.well-known/agents.json
/.well-known/agent-card.json

Security research

Vulnerability reports and disclosures I've submitted through formal channels.

AI Sandbox Trust Boundary Gap — Microsoft Copilot

MSRC VULN-210492 · 2026-09

A Microsoft Copilot bug bounty report closed as "not a vulnerability" that revealed a real gap: the difference between what an AI model refuses to do and what its execution sandbox is actually capable of doing anyway.

Implications

Model-level refusals are a policy layer, not a security boundary — tool execution, sandbox capability, and egress control are separate layers that need independent hardening. Reported to MSRC; disclosed publicly after their review window closed with no objection.

Read the writeup →