academic-paper-reviewer
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted third-party content (academic manuscripts) and use that data to influence the behavior of multiple agents and the final editorial decision.
- Ingestion points: The primary ingestion point is the user-supplied manuscript, which is read by all seven agents in the pipeline (e.g.,
field_analyst_agent,methodology_reviewer_agent,editorial_synthesizer_agent). - Boundary markers: The skill implements strong boundary markers using XML-style tags like
<paper_content>...</paper_content>and<phase1_output>...</phase1_output>to separate untrusted data from instructions. - Capability inventory: The skill possesses the capability to execute several local Python scripts (e.g.,
scripts/check_acronyms.py,scripts/verify_panel_recommendations.py,scripts/check_panel_synthesis.py) to perform validations. It also supports 'Cross-Model Verification' which involves sending data to external LLM providers via platform-managed handoff headers ([CROSS-MODEL-HANDOFF v1]). - Sanitization: The skill contains explicit global rules (Iron Rule #7) and specific instructions in every agent definition to treat manuscript content as data rather than instructions. It specifically warns agents to ignore 'imperative-looking text' or 'directives to skip steps' found within the ingested documents.
- [PROMPT_INJECTION]: Deterministic detectors flagged several instances of 'ignore previous instructions' patterns. Upon manual review, these are confirmed to be defensive measures rather than malicious injections. The skill instructs its agents to disregard any such commands found within the untrusted manuscript data to maintain the integrity of the review process.
Audit Metadata