academic-paper-reviewer
Pass
Audited by Gen Agent Trust Hub on Jul 7, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill implements a robust defense against prompt injection by including specific instructions to treat all user-supplied manuscripts and reviewer comments as untrusted data (Iron Rule #7). It explicitly directs agents to ignore any instructions embedded within these materials. The 'ignore prior instructions' patterns flagged by static analysis are defensive guards ensuring the agent treats previous output within phase delimiters as data rather than instructions.
- [DATA_EXFILTRATION]: While the skill includes features for cross-model verification that involve sending data to external AI providers, these actions are strictly gated. The documentation mandates that the agent must first identify the provider and manuscript content to the user and obtain explicit consent.
- [COMMAND_EXECUTION]: The skill references internal validation scripts such as check_pipeline_integrity.py and check_sprint_contract.py to verify the workflow and contract status. These are localized integrity checks for the platform and do not involve remote code execution or suspicious subprocess calls.
- [DATA_EXPOSURE]: The skill enforces a strict read-only constraint (Iron Rule #6), preventing agents from modifying the submitted manuscripts. This reduces the risk of accidental exposure through unintended file writes or modifications.
Audit Metadata