printing-press-amend
Pass
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill mines Claude Code session transcripts (~/.claude/projects//*.jsonl) to extract 'friction signals'. Since transcripts contain raw history from previous AI turns, they act as an ingestion point for untrusted data that could contain malicious instructions. The skill lacks explicit boundary markers or delimiters to isolate these signals from the agent's control flow. Capabilities include file writing, bash command execution, and performing GitHub PR operations. While a PII scrub is implemented, it does not mitigate prompt injection risks.
- [DATA_EXFILTRATION]: The skill reads sensitive session history stored in ~/.claude/projects/ transcripts. Although the stated purpose is to extract CLI-related friction, this data contains the full context of user interactions. The skill then automates the process of pushing derived content and PRs to a public repository (mvanhorn/printing-press-library).
- [COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool for environment setup, disk space checks, and executing the cli-printing-press binary. It also heavily relies on 'git' and 'gh' CLI tools to manage clones and interact with the GitHub API.
- [EXTERNAL_DOWNLOADS]: During setup, the skill uses 'curl' to fetch version information from raw.githubusercontent.com and prompts the user to install or update the cli-printing-press binary using 'go install' from a GitHub repository.
- [DYNAMIC_EXECUTION]: In Phase 4, the skill autonomously generates code fixes based on identified findings, applies those edits to a local clone, and validates the changes by running 'go build' and other quality gates via the 'publish validate' command.
Audit Metadata