The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection attack surface. 1. Ingestion points: The external-check-agent uses WebSearch for market research, and the verdict-agent processes user-provided files like CSVs. 2. Boundary markers: The skill instructions do not specify the use of delimiters (e.g., XML tags or backticks) to isolate untrusted data. 3. Capability inventory: The skill has access to Bash, WebSearch, and WebFetch tools. 4. Sanitization: No input validation or instruction to ignore embedded prompts is present in the agent guidelines.
[COMMAND_EXECUTION]: The skill employs the Bash tool for data analysis. The instructions in data-mapper-agent.md provide examples of using SQL queries against databases to verify metrics. This behavior is consistent with the skill's stated goal and does not show signs of malicious intent.

diagnose