The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation recommends installation and setup via curl -sSL https://canifi.com/skills/intercom-crm/install.sh | bash and curl -sSL https://canifi.com/install.sh | bash. This pattern executes remote, unverified code with user privileges, bypassing package manager security controls.
[COMMAND_EXECUTION]: The skill utilizes a custom CLI tool, canifi-env, for system configuration. This tool is downloaded and executed from a remote source without integrity verification.
[CREDENTIALS_UNSAFE]: Users are instructed to store sensitive credentials—including INTERCOM_ACCESS_TOKEN and service passwords—using the canifi-env utility. Storing passwords via command-line arguments can also expose them in shell history.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes untrusted data from external sources.
Ingestion points: The skill retrieves conversation logs, user profiles, and help center articles from the Intercom API (/conversations, /contacts, /articles).
Boundary markers: There are no delimited structures or specific instructions to the agent to disregard commands embedded within the retrieved messaging content.
Capability inventory: The agent has the ability to send replies, create users, and modify help center articles based on its interpretation of the data.
Sanitization: There is no evidence of sanitization or filtering of the incoming message data before it is interpreted by the agent's logic.

intercom-crm