The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/gong.sh is vulnerable to shell command injection in the calls and stats commands. Evidence: The variable $DAYS is interpolated into a subshell: FROM=$(date ... || date -d "$DAYS days ago" ...). If an agent passes unsanitized user input here, it could execute arbitrary system commands using shell metacharacters like $(...).
[DATA_EXFILTRATION]: The skill accesses sensitive local credentials. Evidence: SKILL.md and scripts/gong.sh read API keys from ~/.config/gong/credentials.json. While necessary for functionality, this pattern allows sensitive credentials to be ingested into the agent context and potentially exfiltrated.
[COMMAND_EXECUTION]: Potential for JSON injection in API requests. Evidence: In scripts/gong.sh, the transcript and call commands interpolate arguments directly into JSON strings used in curl requests. This could allow a malicious user to modify the API request structure.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface. Ingestion points: Transcripts fetched from the Gong API are processed as raw text in scripts/gong.sh. Boundary markers: None are used to separate transcript content from other instructions. Capability inventory: The skill can read local credential files and perform network requests. Sanitization: No sanitization is applied to the external transcript data before it is presented to the agent.

gong