The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a theoretical attack surface as it processes untrusted data from LinkedIn connection files (CSV/Zip) and external social media post text retrieved via the Crustdata API.
Ingestion points: Local LinkedIn connection exports and live social post content.
Boundary markers: The sub-agent instructions in references/stakeholder-agent.md employ strict JSON formatting requirements and explicit grounding rules to prevent unintended instruction execution.
Capability inventory: The skill executes local Python scripts and uses the Artifact tool to publish HTML files.
Sanitization: The scripts/build_workspace.py script performs HTML entity escaping on all external data before it is injected into the final workspace template, preventing cross-site scripting (XSS) or markdown-based injection in the generated report.
[DYNAMIC_EXECUTION]: The skill utilizes a vendor-provided execute tool from the Crustdata MCP server to run JavaScript snippets. These snippets are used to orchestrate tool calls and handle response projections. This behavior is confined to the specific vendor's execution environment and is necessary for the skill's primary function of data enrichment.
[OBFUSCATION]: The skill uses base64 encoding to embed the vendor's branding assets (logos) and stakeholder photos directly into the generated HTML artifact. This is a standard practice for creating portable, self-contained files and does not involve hiding malicious code or instructions.

warm-path-workspace