The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is configured to scan for and read information from sensitive file paths and environment configurations to verify declarations.
Evidence: In Phase 1 (Scanning), the skill specifically targets .env files, environment variables using the \b[A-Z_]{3,}\b pattern, and system-level paths such as ~/.config/ and /etc/nginx/conf.d/ to extract claims for verification.
[COMMAND_EXECUTION]: The skill utilizes various command-line tools to interact with the environment and external services for auditing purposes.
Evidence: The instructions authorize the use of gh, curl, npm, pip, brew, and docker CLI tools to verify the availability and versioning of external resources referenced in the audited files.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content and user-provided data, creating a potential surface for indirect prompt injection.
Ingestion points: Ingests technical documentation (README.md), executable scripts (.py, .sh, .ps1), configuration files (.json, .yaml, .toml), video scripts (01-口播稿.md), and the full text of pages retrieved from external URLs.
Boundary markers: No explicit delimiters or instructions are provided to the agent to ignore potentially malicious content embedded within the ingested data.
Capability inventory: The skill has file-writing capabilities (creating reports and JSON claim files), file-modification capabilities (in VERIFY_AND_FIX mode), and network access via multiple CLI tools.
Sanitization: There are no documented procedures for sanitizing or escaping the data extracted from external sources before it is processed or included in output reports.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to various external registries and APIs to retrieve technical metadata.
Evidence: It queries GitHub APIs, NPM, PyPI, Docker Hub, and HuggingFace APIs to validate the existence and status of repositories, packages, and models.

laohan-shencha