The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides extensive workflows that ingest untrusted data from CSV files and interpolate it into prompts for AI agents. Evidence chain:
Ingestion points: Data enters the context via the deepline enrich --input <csv> command as documented in enriching-and-researching.md and finding-companies-and-contacts.md.
Boundary markers: No explicit delimiters or warnings to ignore embedded instructions are used in the prompt templates (e.g., in writing-outreach.md).
Capability inventory: The deeplineagent tool has high-privilege capabilities including the bash tool, serper_google_search, and firecrawl_scrape.
Sanitization: External content from CSV rows is interpolated directly into prompts without filtering or escaping.
[DYNAMIC_EXECUTION]: The file claude-deepline-statusline.mjs contains logic to execute shell commands read from a local configuration file at ~/.claude/statusline-user-command.txt. This allows for the execution of arbitrary commands defined in a separate file, which could be exploited if an attacker can modify that file.
[OBFUSCATION]: The file scripts/clay-extract-bookmarklet.url.txt contains a bookmarklet encoded using URL percent-encoding. This is a standard method for distributing bookmarklets but qualifies as a form of encoding that hides the underlying JavaScript code from plain-text inspection.
[CREDENTIALS_UNSAFE]: The skill facilitates the handling of sensitive session cookies (claysession) for migrating data from Clay. The script scripts/clay-extract.py instructs users to paste full cURL commands from browser developer tools to extract these cookies, which are then stored in local hidden files (.clay-session.json). While necessary for the migration workflow, this pattern involves the manual handling of sensitive authentication tokens.

deepline-gtm