The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous curl commands intended for network probing. These commands are used to identify VPN appliance versions and test for specific historical vulnerabilities such as path traversal and memory disclosure.
[INDIRECT_PROMPT_INJECTION]: The agent is instructed to fetch and process responses from external, potentially untrusted VPN servers, which presents a surface for indirect prompt injection.
Ingestion points: External HTTP/HTTPS responses from targets fetched via curl in SKILL.md.
Boundary markers: The skill includes a 'Mandatory confirmation gate' requiring explicit user authorization, scope verification, and command preview before any active probing.
Capability inventory: Extensive network probing capability via curl, file system read capability (in the context of traversal testing), and local payload generation via python3 across the entire SKILL.md file.
Sanitization: No automated sanitization or filtering of remote responses is implemented; the agent relies on its own internal safety policies and user confirmation.
[CREDENTIALS_UNSAFE]: The file contains a reference table listing well-known factory default credentials for several enterprise VPN vendors (e.g., Cisco ASA, Citrix NetScaler, Palo Alto) to assist in authorized security testing.
[DYNAMIC_EXECUTION]: Employs a Python one-liner (python3 -c "print('A' * 24812)") to dynamically generate a large string payload required for testing the 'Citrix Bleed' (CVE-2023-4966) vulnerability.

enterprise-vpn-attack