The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides extensive documentation for executing shell commands through the bash and remote-bash step types, which are core components of the security scanning workflows.
[PRIVILEGE_ESCALATION]: Documentation includes mechanisms for handling elevated permissions, such as the sudo_auth utility function and the --sudo-aware CLI flag, intended for tasks requiring root access during security audits.
[INDIRECT_PROMPT_INJECTION]: The guide details the use of agent and agent-acp steps which process external scan data (such as HTTP headers and target output) using LLMs with tool access. While this exposes an attack surface for indirect prompt injection, it is a documented feature for automated security analysis.
Ingestion points: Agent queries in references/examples.md and references/step-types.md process external data like {{live_hosts}} and HTTP response content.
Boundary markers: Explicit delimiters for untrusted data are not mandated in the basic documentation examples.
Capability inventory: Agent tools include bash, read_file, save_content, and http_get across various scripts.
Sanitization: Examples do not specifically demonstrate input sanitization for data passed to LLM agents.
[DYNAMIC_EXECUTION]: The Osmedeus engine uses the Goja JS runtime for evaluating workflow expressions and scripts, and supports execution of inline Python (exec_python) and TypeScript (exec_ts) code.
[EXTERNAL_DOWNLOADS]: The skill documents capabilities to fetch remote workflow modules via the --module-url flag and to install base components and tool binaries using the osmedeus install command.

osmedeus-expert