manuscript-scaffold

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFECREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSPERSISTENCEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill handles sensitive Git authentication tokens for Overleaf integration.
  • Evidence: Phase 3 describes collecting a token and storing it using git credential-store or within ~/.netrc.
  • Mitigation: The skill explicitly instructs to never write the token to tracked files or logs and requires masking the token in all outputs except for the last four characters.
  • [PERSISTENCE]: The skill establishes persistent credentials for Git operations.
  • Evidence: Phase 3, Branch B, step B4 modifies ~/.git-credentials or ~/.netrc to facilitate ongoing authentication with git.overleaf.com.
  • Mitigation: This is a standard Git credential management practice and requires explicit user choice between storage methods.
  • [EXTERNAL_DOWNLOADS]: The skill allows downloading LaTeX templates and cloning remote repositories.
  • Evidence: Phase 2 permits downloading journal archives from user-provided URLs; Phase 3 performs git clone from Overleaf.
  • Mitigation: Downloads from external URLs are gated by a mandatory SHA256 hash display and explicit user confirmation before extraction. Overleaf is recognized as a well-known service for research workflows.
  • [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface where instructions could be embedded in the project metadata it processes.
  • Ingestion points: Data is ingested from AGENTS.md (fields like MANUSCRIPT_DIR, TARGET_VENUE, and OVERLEAF_GIT_URL) and direct user input (tokens and URLs).
  • Boundary markers: The skill does not define specific prompt delimiters but relies on multi-phase execution and explicit user confirmation prompts for all risky actions.
  • Capability inventory: The skill can perform file system modifications (mkdir, cp), execute shell commands (git, awk, tlmgr), and perform network operations (git clone/push, curl/wget).
  • Sanitization: Includes URL format validation for Overleaf links and mandatory SHA256 integrity checks for external archive downloads.
  • [COMMAND_EXECUTION]: The skill executes various system commands to manage the research environment.
  • Evidence: Uses git for version control, tlmgr for LaTeX package management, and shell utilities like awk for parsing branch information.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 09:21 AM
Security Audit — agent-trust-hub — manuscript-scaffold