The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive workspace, usage, and billing information from the DCE platform via the dce CLI. This data is collected into a temporary directory to perform attribution analysis. This access is authorized for the skill's purpose and does not involve exfiltration to external domains.
[COMMAND_EXECUTION]: The skill executes a bundled shell script scripts/collect_margin_attribution.sh which orchestrates multiple calls to the dce CLI. The script correctly quotes variables derived from user input and incorporates basic sanitization for usernames used in generated filenames, minimizing the risk of command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes JSON data from external APIs (DCE/Crane) and user-provided cost files, creating a surface for indirect prompt injection. While the skill lacks explicit boundary markers for this data, it applies a strict mathematical framework for attribution which reduces the impact of potentially malicious content in the data source.
Ingestion points: API responses from dce and user-provided model-cost JSON files in SKILL.md and scripts/collect_margin_attribution.sh.
Boundary markers: None defined for the interpolation of external data into the agent's logic.
Capability inventory: The agent has the capability to execute shell commands through the bundled script and the dce platform tool.
Sanitization: The shell script uses sed to sanitize username inputs for file path construction.

llm-gross-margin-attribution