The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to clone a repository from an unverified GitHub account (vietjovi/pentest-ai-killer) and execute scripts and build binaries from it.
[PRIVILEGE_ESCALATION]: Manual installation instructions explicitly require running a shell script with root privileges via 'sudo bash install_pentest_tools.sh', which can execute arbitrary code on the host with elevated permissions.
[REMOTE_CODE_EXECUTION]: The tool setup involves downloading, compiling, and executing remote code from an untrusted external source, including a Go-based CLI and multiple third-party security binaries.
[COMMAND_EXECUTION]: The REST API and MCP bridge expose a 'custom_command' parameter (e.g., in the /tools/nuclei endpoint) that allows raw shell commands to be passed directly to the host's execution environment, enabling trivial command injection.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted target data (URLs, domains, cloud profiles) and processes them through various subprocess calls, creating an injection surface.
Ingestion points: 'targets' parameter in API calls and natural language prompts for scanning targets.
Boundary markers: None provided in the documentation or MCP setup to prevent the AI from being manipulated into executing malicious commands.
Capability inventory: The skill has extensive capabilities including full network access, file system read/write for reports, and subprocess execution for 69+ integrated tools.
Sanitization: No evidence of input validation or sanitization is mentioned for the 'custom_command' or target inputs.

pentest-ai-killer-automation