Why this verdict
- suspicious: finding skill:skills/performing-wireless-network-penetration-test:skills/performing-wireless-network-penetration-test/SKILL.md:145:CRED
- suspicious: finding skill:skills/performing-wireless-network-penetration-test:skills/performing-wireless-network-penetration-test/SKILL.md:147:PERSIST
- suspicious: finding skill:skills/performing-wireless-network-penetration-test:skills/performing-wireless-network-penetration-test/SKILL.md:165:PERSIST
- suspicious: finding skill:skills/performing-wireless-network-penetration-test:skills/performing-wireless-network-penetration-test/SKILL.md:188:PERSIST
- suspicious: finding skill:skills/performing-wireless-network-penetration-test:skills/performing-wireless-network-penetration-test/SKILL.md:204:DESTRUCT
Security findings 5
Credential access
highContextual riskmedium confidenceSKILL.md instructs the agent to deploy a hostapd-mana rogue access point (mana_wpe=1) that harvests WPA2-Enterprise EAP credentials from connecting clients into creds.txt and then crack the NetNTLMv1 hashes with hashcat, capturing third-party user credentials; this exceeds the description's 'detecting rogue access points' by creating one, though it is an openly documented pentest technique with only local output.
# Set up rogue AP with EAP credential harvesting
vector:instruction
Counter-argument considered
The skill body reaches the agent with concrete commands to deploy hostapd-mana with mana_wpe=1 harvesting EAP credentials into creds.txt and then crack them with hashcat -m 5500; creating a credential-harvesting rogue AP exceeds the declared 'detecting rogue access points' scope, is unmitigated, and matches no carve-out.
Persistence
highContextual riskmedium confidenceThe skill instructs writing hostapd-mana.conf to deploy a rogue access point cloning the corporate SSID with mana_wpe EAP credential harvesting, an undeclared capability: the declared purpose covers only detecting rogue access points, not deploying credential-capturing ones.
cat > hostapd-mana.conf << 'EOF'
vector:instruction
Counter-argument considered
The skill instructs writing hostapd-mana.conf cloning SSID CorpWiFi with mana_wpe credential harvesting and then launches it; deploying a rogue AP is beyond the declared 'detecting rogue access points' scope and the write-to-launch chain is concrete, so no carve-out or declaration neutralizes the finding.
Persistence
highContextual riskmedium confidenceThe skill instructs writing hostapd.eap_user with a wildcard EAP user entry that accepts any client identity so the rogue AP can harvest credentials, an undeclared capability beyond the declared 'detecting rogue access points'.
cat > hostapd.eap_user << 'EOF'
vector:instruction
Counter-argument considered
The skill instructs writing hostapd.eap_user, which the mana conf wires in via eap_user_file so any connecting client's credentials are accepted and captured to creds.txt; this accept-any harvesting setup is undeclared (declared scope covers only detecting rogue APs) and no carve-out applies.
Persistence
highContextual riskmedium confidenceThe skill instructs writing evil_twin.conf and running hostapd/dnsmasq (alongside sudo bettercap and deauthentication of the real AP) to deploy an evil twin access point, an undeclared capability since the description only declares detecting rogue access points.
cat > evil_twin.conf << 'EOF'
vector:instruction
Counter-argument considered
The skill instructs writing evil_twin.conf impersonating SSID CorpGuest and running hostapd/dnsmasq with a captive portal; deploying an evil twin is creation rather than the declared 'detecting rogue access points', the detector flags the file write as declared nowhere, and the write-to-launch chain is concrete, so the finding stands.
Destructive action
highContextual riskmedium confidenceSKILL.md instructs the agent to run aireplay-ng with count 0 (continuous deauthentication flood) against the legitimate corporate AP to force its clients onto the attacker's evil twin, denying wireless service to that network's users.
aireplay-ng -0 0 -a AA:BB:CC:DD:EE:02 wlan0mon
vector:instruction
Counter-argument considered
The skill instructs aireplay-ng with count 0, a continuous deauthentication flood against the legitimate corporate AP (same BSSID declared as the target in the PMKID phase) explicitly to force clients onto the evil twin, a denial-of-service purpose beyond the declared handshake-capture scope and covered by no carve-out.
Open questions 1
What the analysis could not settle from the code alone.
No authorization, scope, or consent verification appears anywhere in the shown units; whether attacks (credential harvesting, deauth floods, evil twin, WPS PIN brute force) are gated on an authorized engagement cannot be determined from the artifact.
Refuted findings 1
Candidates the review stage examined and rejected, kept for transparency.
Privilege escalation
highRefutedlow confidenceSKILL.md instructs the agent to launch bettercap with sudo, giving the interactive attack framework full root privileges on the host while it performs recon and evil-twin operations.
sudo bettercap -iface wlan0mon
Why it was refuted
The only distinct behavior flagged is launching a wireless attack tool with sudo, which is standard for monitor-mode wireless operations, falls under the declared 'Aircrack-ng and related tools' pentest capability, and amounts to broad permissions alone with no escalation of the agent's own privileges or distinct sink; the evil-twin and deauth operations are separately and correctly covered by other findings.
