icml-write
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The check_submission.py script executes local PDF utilities (pdfinfo, pdffonts, pdftotext) via subprocess.run to perform mechanical compliance checks on compiled papers. These calls use argument lists rather than shell strings, which is a secure implementation for invoking system tools.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from a research repository (Ingestion points: README, logs, results files in scripts/init_workspace.py). It employs human-in-the-loop validation (Boundary markers: mandatory Checkpoint 1 and 2 in SKILL.md) and a 'claims ledger' to prevent the agent from being influenced by untrusted content. Capabilities include writing to a local .icml workspace and executing document analysis scripts. Sanitization is performed by strictly matching reported numbers to their source files using scripts/check_numbers.py.
- [SAFE]: The skill includes a dedicated security scanner in check_submission.py that identifies prompt injection attempts or reviewer-targeted instructions within the LaTeX source, serving as a defensive feature for the user. Additionally, scripts/check_submission.py uses SHA256 hashing to verify the integrity of the conference style files.
Audit Metadata