ecc-conventions
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The 'Cross-harness skill contribution review workflow' processes external data from a JSON file (
contribution-review-observation.json) which is then used by a script to generate a review 'Context Capsule'. This data ingestion creates an attack surface where instructions embedded in the JSON could influence the agent's downstream behavior. - Ingestion points: The workflow in
SKILL.mdingest facts and evidence fromcontribution-review-observation.jsoninto thecompile_contribution_review.pyscript. - Boundary markers: The instructions do not specify any delimiters or warnings to ignore instructions embedded within the observation data.
- Capability inventory: The skill utilizes shell command execution (directory creation, file copying), Python script execution, and unit test execution.
- Sanitization: While the skill claims to check structure and consistency, it does not explicitly describe sanitization or escaping of the interpolated content to prevent prompt injection.
- [COMMAND_EXECUTION]: The skill provides explicit shell command templates for the agent to execute, including file system operations (
mkdir,cp) and execution of local Python tools and test runners.
Audit Metadata