The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requests the use of dangerouslyDisableSandbox: true when performing network scans through tunnels, which removes security protections from the execution environment to facilitate specific networking tools.
[EXTERNAL_DOWNLOADS]: Instructions guide the user or agent to download static binaries from an external GitHub repository (github.com/andrew-d/static-binaries) which is not an officially verified or trusted organization repository.
[REMOTE_CODE_EXECUTION]: The skill provides instructions for transferring and executing these external binaries on pivot hosts using methods like Base64 encoding and local execution.
[COMMAND_EXECUTION]: The skill uses extensive shell commands and complex script blocks (Bash and PowerShell) to perform network discovery and host enumeration.
[DATA_EXFILTRATION]: The skill is designed to discover and save detailed network infrastructure data, including sensitive information like service banners and evidence of vulnerabilities, to local directories.
[PROMPT_INJECTION]: Keywords associated with malicious behavior, such as 'reverse shell', were detected in the instruction text. While used contextually to describe legitimate penetration testing scenarios, they match patterns used for behavior override attempts.
[INDIRECT_PROMPT_INJECTION]: The skill parses structured and unstructured output from network tools (e.g., nmap XML).
Ingestion points: Reads output from nmap, grep, and awk commands.
Boundary markers: Instructions request output in specific formats but do not implement delimiters to prevent the agent from following instructions embedded in tool results.
Capability inventory: Significant shell execution capabilities are available to process the ingested data.
Sanitization: There is no evidence of sanitization or validation of hostnames, service names, or banners returned by the network tools before they are processed by the agent.

network-recon