The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external TrueFoundry API responses, such as resource names, IDs, and descriptions. This data is interpolated into subsequent prompts and command payloads, creating a surface where malicious content in the platform's metadata could influence agent behavior.
Ingestion points: Data enters the context via tfy-api.sh output from resource listing endpoints in SKILL.md and references/access-control.md (e.g., workspace names, cluster IDs, and team descriptions).
Boundary markers: Present. The instructions include explicit directives to confirm resources and subjects with the user before performing actions, such as "Never auto-pick a workspace" and "Confirm subject, role, and resource with the user."
Capability inventory: The skill uses the Bash tool to perform network operations (curl) and can call tools to modify access controls, manage secret groups, and generate tokens.
Sanitization: Absent. There are no specific instructions for escaping or validating the string content of the API responses before they are used in the agent's context.
[COMMAND_EXECUTION]: The skill executes shell commands via a local script tfy-api.sh to interact with the TrueFoundry API. While these operations are necessary for the skill's functionality, they involve sending payloads (JSON strings) constructed from agent context and user input. The skill mitigates potential command injection and unintended action risks by requiring human-in-the-loop (HITL) confirmation for all creation, update, and access granting operations.

truefoundry-platform