The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It is designed to ingest and analyze untrusted external data such as team composition, role descriptions, and offboarding plans. An attacker could embed malicious instructions within these role definitions to influence the agent's output or scoring logic. Evidence: The skill gathers context in Step 0 and scores roles in Step 1 based on external inputs. There are no explicit boundary markers or sanitization steps mentioned to separate data from instructions.
[COMMAND_EXECUTION]: The skill requires the Bash tool to perform its audit functions. This allows for the execution of shell commands on the local system. While intended for reading files and processing data, this capability could be exploited if the agent is tricked into executing unintended commands during the planning phase.
[DATA_EXFILTRATION]: The skill's primary function involves handling highly sensitive organizational information, including employee roles, headcount, and severance frameworks. While no network-based exfiltration tools are currently permitted, the use of Bash and Read enables the potential for sensitive data to be moved or exposed within the local environment.

folk-migrate