The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture handles untrusted external data which could lead to manipulation of agent instructions.\n
Ingestion points: Data enters the context via the web_fetch tool when analyzing landing pages (referenced in references/linkedin.md) and through user-provided account exports, screenshots, and CRM data during the audit mode.\n
Boundary markers: The instructions in SKILL.md and the reference playbooks do not contain explicit delimiters or 'ignore' instructions to separate processed data from command execution context.\n
Capability inventory: The skill is configured in agents/openai.yaml with broad tool access including bash, run_command, write_to_file, replace_file_content, and web_fetch.\n
Sanitization: There are no documented procedures for sanitizing or validating external content before it is ingested and acted upon by the agent.\n- [DATA_EXFILTRATION]: The skill is designed to manage sensitive business and marketing data.\n
Sensitive data handling: The skill interacts with campaign budgets, lead-generation data (PII), and conversion tracking metrics across multiple platforms.\n
Exfiltration risk: The combination of access to sensitive account information and the ability to perform network operations via web_fetch creates a theoretical risk for data exfiltration if the agent is compromised by malicious external content.

paidads