The Agent Skills Directory

[PROMPT_INJECTION]: The skill describes building agents that ingest untrusted user input (e.g., customer IDs or issue descriptions) which is then interpolated into FEEL prompt strings. This creates a surface for indirect prompt injection where an external actor might attempt to influence the agent's tool-calling behavior. This is an inherent risk of AI agent architectures and is documented here as part of the primary functionality of the skill.
Ingestion points: Untrusted data enters the agent context via variables (e.g., customerId, issue) defined in SKILL.md and used in prompt templates.
Boundary markers: The examples show the use of FEEL expressions (starting with =), but do not explicitly detail the use of delimiters to separate user input from system instructions.
Capability inventory: The agents have capabilities including making network requests via REST connectors, executing FEEL expressions in script tasks, and interacting with human users, as described in references/tool-modeling.md.
Sanitization: The provided examples do not demonstrate explicit sanitization or validation of the input variables prior to prompt interpolation.
[DATA_EXFILTRATION]: No exfiltration patterns were identified. The skill proactively guides users to manage credentials safely by using Camunda cluster secrets ({{secrets.*}}) and explicitly warns against hardcoding sensitive information.
[EXTERNAL_DOWNLOADS]: The skill references documentation and command-line tools (such as c8ctl) associated with the official Camunda domain (docs.camunda.io). These are verified as trusted vendor resources.

camunda-ai-agent