<< All versions

Skill v1.0.0

currentAutomated scan100/100
tobiasveiga00/claude-security-audit/network
──Details
PublishedSeptember 27, 2026 at 09:28 PM
Content Hashsha256:ef0d8a63b77e97fc...
Git SHA
──Files
Files (1 file, 4.8 KB)
SKILL.md4.8 KBactive
SKILL.md · 131 lines · 4.8 KB

version: "1.0.0" name: network description: Network and wireless security assessment — host and service discovery, TLS posture, exposed services, and Wi-Fi testing (WPA2/WPA3, PMKID, 802.1X EAP, evil twin, PMF). All active work is gated behind written rules of engagement. Use for "network scan", "port scan", "check my TLS", "Wi-Fi audit", "wireless assessment". argument-hint: "[host, range, or --wireless]" allowed-tools: Read, Write, AskUserQuestion, Bash(node:), Bash(nmap:), Bash(testssl.sh:), Bash(naabu:), Bash(hcxdumptool:), Bash(hashcat:), Bash(hostapd-wpe:*)


Network and wireless security

Everything in this skill puts packets on a wire. None of it runs until the authorization gate clears it. That is not a formality — an unauthorized scan is, in many jurisdictions, a criminal act.

Rules of engagement: !node "${CLAUDE_PLUGIN_ROOT}/scripts/scope.mjs" show 2>&1

Tooling: !node "${CLAUDE_PLUGIN_ROOT}/scripts/tools.mjs" --domain network,wireless --missing 2>&1


Before anything

If there is no scope, stop and run /security-audit:scope. If there is, confirm the specific target is covered:

bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/scope.mjs" check "nmap -sV target"

The gate will block an out-of-scope target regardless, but checking first lets you fix the command instead of hitting a wall.


Network

Discovery

Start narrow. A full scan of a large range is slow, noisy, and rarely what the engagement needs first.

bash
nmap -sV -sC -oX .security-audit/nmap.xml <in-scope target>

Then import the results rather than reading the XML into the conversation:

bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/finding.mjs" import --tool sarif --file <converted>

(Nmap XML has no importer; summarise the interesting services and record findings by hand, or convert to SARIF first.)

What to look for

  • Unexpected exposed services. Databases, admin panels, message brokers,

Kubernetes API, Docker daemon, Redis, Elasticsearch — anything data-bearing that answers from the internet.

  • Default or weak service configuration. Anonymous FTP, open SMB shares,

SNMP with public, unauthenticated Redis.

  • Version-based exposure. Cross-reference banners against known CVEs:

node "${CLAUDE_PLUGIN_ROOT}/scripts/intel.mjs" cve <id>.

TLS

bash
testssl.sh --jsonfile .security-audit/testssl.json <host:port>

Report protocol versions (SSLv3, TLS 1.0/1.1 should be gone), weak cipher suites, certificate validity and chain, Secure renegotiation, and known issues (Heartbleed, ROBOT) where the tool flags them.


Wireless

Wireless is different in kind: radio attacks reach every device in range, not only your target. The gate therefore requires SSIDs or BSSIDs to be declared explicitly, and every one on a command line must match a declared target.

Weight the assessment toward what actually matters in 2026 (full rationale in ${CLAUDE_PLUGIN_ROOT}/references/frameworks.md):

High priority

  • WPA3 transition-mode downgrade — the number-one WPA3 finding. Transition

mode keeps a WPA2 path; force it, capture, crack offline. The control is SAE-only with transition mode disabled.

  • 802.1X / EAP misconfiguration — the top enterprise finding. Absent server

certificate validation, no CA pinning, no ServerName constraint, crackable MSCHAPv2. Test with a controlled hostapd-wpe, never against a live corporate network without explicit sign-off.

  • PMKID capture — clientless. hcxdumptool to capture, hashcat -m 22000

to crack. The fastest WPA2-PSK passphrase audit.

  • Evil twin / rogue AP — unaffected by WPA3 for open and captive-portal

cases. The primary credential-harvest vector.

Medium priority

  • WPS Pixie Dust / PIN brute force — still enabled on plenty of SOHO gear.

The finding is "WPS enabled".

  • 4-way handshake capture — meaningful against PSK only, never SAE.
  • Karma / PNL abuse — client-side; persists on legacy and provisioned

profiles.

As a control, not an attack

  • 802.11w PMF state — mandatory under WPA3, optional under WPA2. Audit the

state; deauthentication is only the symptom.

  • KRACK and Dragonblood — patched across mainstream stacks. Report as

firmware-currency checks; the residual surface is unmanaged IoT and legacy embedded.

Deauthentication is a denial of service

It needs denialOfService: true in scope, and it affects bystanders on shared spectrum. Confirm the client genuinely wants it, in writing, before enabling it.


Recording

domain: "network" or "wireless", the host and port or the SSID/BSSID in location, and evidence that is a real captured artefact, redacted. For wireless, name the specific configuration control (SAE-only, PMF required, WPS disabled) in the remediation.

All versions