The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the AI agent to execute a local Python script using command-line arguments populated from user input, specifically for domains to be banned and names of enrichment tools.\n
Evidence: SKILL.md contains the instruction: run scripts/account_list_check.py --list list.txt --research research.csv --ban <domains> [--enrichment <tool names>].\n
Risk: If the agent interpolates these user-supplied values into a shell command string without escaping or sanitization, a malicious actor could inject additional commands using shell metacharacters (e.g., semicolons, pipes, or backticks).\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from various sources (web pages, CRM files, and pasted text) that could contain malicious instructions designed to hijack the agent.\n
Ingestion points: SKILL.md (Step 1, Step 3, and Step 8) specifies reading data from fetched pages, uploaded files, and third-party text.\n
Boundary markers: The skill contains explicit instructions in its preamble and 'Golden Rule' to treat all external inputs as data rather than instructions. It requires a mandatory notification to the user if such instructions are encountered.\n
Capability inventory: The agent is authorized to use scripts/account_list_check.py and external CRM/enrichment data connectors.\n
Sanitization: The skill mandates reporting detected injection attempts in the final brief and explicitly ignoring the malicious content, establishing a behavioral sanitization pattern.

account-research