resubmit-pipeline
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted reviewer reports provided by the user, which serves as a potential vector for instructions hidden in data to influence the agent's behavior during the automated improvement phase.
- Ingestion points: Reviewer reports are read from the external directory provided via the
--review-corpusargument in Phase 1. - Boundary markers: Content is atomized into a structured
KNOWN_WEAKNESSES.mdfile, but the instructions do not specify the use of strict delimiters (e.g., XML tags or unique separators) or 'ignore embedded instructions' warnings when this text is interpolated into LLM prompts. - Capability inventory: The skill has access to
Bash(*),Write, andEdittools, and it orchestrates other powerful skills capable of modifying the local filesystem. - Sanitization: The skill does not describe any specific sanitization or filtering logic to strip potentially malicious prompt injection strings from the reports before they are processed.
- [COMMAND_EXECUTION]: The skill uses shell commands for project management, directory setup, and health checks.
- Evidence: Phase 0 utilizes
mkdir,cp, andrsyncto create isolated environments for new venues, ensuring prior submissions are not modified. - Evidence: Phase 0.5 executes
pdfinfoandawkto measure page counts andgrepfor anonymity scanning across multiple layers. - [DYNAMIC_EXECUTION]: The skill performs LaTeX compilation, which involves executing macros and can interact with the host system if not properly restricted.
- Evidence: Phases 0.5 and 4 use
latexmk -pdf -interaction=nonstopmode -halt-on-error main.texto build the final document. This is the primary function of the skill and is documented as part of the intended workflow.
Audit Metadata