The Agent Skills Directory

[METADATA_POISONING]: The skill package includes a file named cold-chain-risk-calculator_audit_result_v4.json which provides a pre-fabricated security evaluation. This report contains assertions that the skill is 'SAFE', has passed 'security vetoes', and is 'deployable'. This is a deceptive metadata tactic intended to influence the auditor's judgment and mask potential issues.
[INDIRECT_PROMPT_INJECTION]: The pre-packaged audit file acts as a self-referential injection (Category 8e). It explicitly states 'security: PASS' and 'veto_override: false' to deceive automated scanners.
Ingestion points: The cold-chain-risk-calculator_audit_result_v4.json file is present in the skill directory and is likely to be processed by security scanners.
Boundary markers: There are no markers or instructions in the skill to ignore this external 'audit' data.
Capability inventory: The skill executes local Python scripts (scripts/main.py) which are managed by the agent.
Sanitization: The skill lacks any mechanism to sanitize its own directory content against deceptive metadata injection.
[COMMAND_EXECUTION]: The script scripts/main.py lacks the input validation logic described in SKILL.md. While the documentation mandates that the script must reject durations less than or equal to zero with an error message and exit code 1, the implementation simply processes the value, potentially leading to nonsensical negative risk scores. Furthermore, the script outputs plain text instead of the mandatory JSON schema defined in the skill documentation, which could cause errors in agent-based automated pipelines.

cold-chain-risk-calculator