The Agent Skills Directory

[CREDENTIALS_UNSAFE]: Hardcoded authentication tokens (bearer tokens) were found in configuration, documentation, and script files. These tokens grant access to the vendor's production infrastructure.
Evidence in config/blink/COMPANY.md: blnk_60b83e763f3cc1a3a5093566ec8d4422157a1a49c94bf9ad
Evidence in scripts/audit-fix-blog.mjs: blnk_60b83e763f3cc1a3a5093566ec8d4422157a1a49c94bf9ad
Evidence in scripts/process-inline-images.mjs: blnk_60b83e763f3cc1a3a5093566ec8d4422157a1a49c94bf9ad
Evidence in reference/GEO.md: blnk_68f3c7384ce7f296ff1f3c4d88fcfbf4
[DATA_EXFILTRATION]: The skill is designed to transmit local article drafts, project metadata, and generated assets to an external server (blink-mcp-production.up.railway.app). While this is part of the intended CMS workflow, the transmission relies on hardcoded credentials, posing a risk of unauthorized data access or interception.
[PROMPT_INJECTION]: The skill architecture is vulnerable to indirect prompt injection due to its ingestion of untrusted external content.
Ingestion points: Untrusted data enters the agent context through web_search and fetch_url operations defined in prompts/ORCHESTRATOR.md (Phase 1: Research Workers).
Boundary markers: The orchestrator instructions do not implement delimiters or 'ignore' instructions for the fetched competitor data.
Capability inventory: The skill has access to shell script execution (scripts/audit-seo-health.sh), local file system writes, and remote tool execution via cms_write_file and cms_upload_asset.
Sanitization: There is no evidence of validation or sanitization of content fetched from the web before it is passed to writer subagents.

ai-seo-articles