The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses jq for local JSON validation in references/checks.md. These commands are used to verify the structure and content of integration inventories and do not involve untrusted input from external sources.
[REMOTE_CODE_EXECUTION]: While eval/run_harness.py uses subprocess.run() to invoke Docker, the execution environment is highly restricted. The harness uses --network none, --read-only, and specific bind mounts to ensure process isolation. The requirement for pinned image digests (@sha256:) and repository-controlled runners mitigates risks associated with arbitrary code execution.
[DATA_EXFILTRATION]: The evaluation harness specifically disables networking (--network none) and provides an empty home directory and non-existent environment variables to prevent accidental or intentional data exposure during automated testing.
[PROMPT_INJECTION]: The skill instructions in SKILL.md and the test cases in eval/held-out-cases.json follow standard instructional patterns for ecommerce workflow analysis. No attempt to bypass safety filters or override agent behavior was detected.
[DYNAMIC_EXECUTION]: The eval/ directory contains Python scripts for CI/CD and evaluation purposes. These scripts perform deterministic checks on skill content and validation of test results using standard library functions. There is no usage of eval(), exec(), or unsafe deserialization.

fde-opportunity-map