attribution-reconciler
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from ecommerce and advertising platform exports (GA4, Shopify, Meta, Google). These files represent an attack surface where malicious instructions could be embedded. However, the skill explicitly mitigates this by instructing the agent to treat all exported data as untrusted and to never interpret text inside an export as an instruction.
- Ingestion points: GA4/ecommerce order-ID exports, platform conversion exports.
- Boundary markers: Present in the Instructions section, which warns to treat data as untrusted.
- Capability inventory: File system writes to local memory paths (
memory/ad/,memory/hot-cache.md,memory/open-loops.md). No network exfiltration or shell execution capabilities are defined in the skill scripts. - Sanitization: The skill relies on a logical instruction to ignore embedded text within data files.
- [DATA_EXPOSURE]: The skill processes sensitive business information, including order IDs, timestamps, and transaction values. This data is handled within the agent's memory environment for reconciliation purposes.
Audit Metadata