The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch and process data from various external sources that are often under the control of an adversary in a CTF or real-world scenario.
Ingestion points: External data enters the context through curl commands (e.g., BlueSky API, IP geolocation services, Wayback Machine), dig output (DNS TXT records), nmap or nc banners, and social media post parsing (e.g., Unicode homoglyphs).
Boundary markers: The instructions lack explicit boundary markers or warnings to the agent to treat data from these external sources as untrusted content that might contain embedded instructions.
Capability inventory: The agent is provided with extensive capabilities including broad shell access (Bash Read Write Edit Glob Grep), the ability to install new software via pip, npx, apt, and brew, and file system access.
Sanitization: There is no evidence of sanitization, escaping, or filtering of the external content before it is processed by the agent or interpolated into its operations.
[COMMAND_EXECUTION]: The skill requires the execution of system-level commands to set up the environment and perform analysis.
Evidence: The SKILL.md file contains instructions to execute apt install, brew install, and pip install to add numerous third-party utilities and libraries to the system.
Evidence: tools-mcp.md includes the command npx @playwright/mcp@latest install-browser chrome-for-testing, which downloads and installs a browser binary to the local environment.
[EXTERNAL_DOWNLOADS]: The skill frequently fetches content and tools from external sources.
Evidence: Fetches browser automation tools from Microsoft's official GitHub repository via npx @playwright/mcp.
Evidence: Clones external codebases such as WebBreacher/WhatsMyName from GitHub to perform username enumeration.
Evidence: Makes numerous network calls to third-party APIs for geolocation (ip-api.com), archiving (web.archive.org), and social media metadata (public.api.bsky.app).

ctf-osint