The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes meeting transcripts and summaries which are external, potentially untrusted sources of data. This creates a surface for Indirect Prompt Injection if the meeting content contains malicious instructions intended to manipulate the AI agent.
Ingestion points: scripts/get-transcript.sh and scripts/get-summary.sh fetch meeting content from the Fathom API.
Boundary markers: Absent. The scripts output the raw transcript and summary text directly to stdout without using delimiters or safety instructions to separate the data from agent instructions.
Capability inventory: The skill has the ability to execute shell commands (curl, jq), access the network (api.fathom.ai), and read local configuration files (~/.fathom_api_key).
Sanitization: Absent. The meeting content is extracted via jq and printed directly without filtering or escaping.
[COMMAND_EXECUTION]: Shell scripts in the scripts/ directory construct API requests and payloads using string interpolation of variables that are not strictly validated.
Evidence: In scripts/setup-webhook.sh, the $WEBHOOK_URL variable is inserted into a JSON payload string. If the input contains escaped double-quotes or control characters, it could be used to manipulate the JSON structure sent to the API.
Evidence: In scripts/list-calls.sh and scripts/search-calls.sh, query parameters like $LIMIT and $CURSOR are appended directly to the URL string. While used within curl calls, a lack of URL-encoding or character validation could lead to unexpected behavior if special characters are provided as input.

fathom