asciinema-analyzer

Warn

Audited by Gen Agent Trust Hub on Oct 17, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill includes 'Self-Evolving' instructions that command the agent to modify the main SKILL.md file immediately if issues are encountered. This creates a loop where the agent's core logic can be altered by runtime experiences or untrusted data processed by the skill.
  • Evidence in SKILL.md: 'Self-Evolving Skill: This skill improves through use. If instructions are wrong... fix this file immediately, don't defer.'
  • Evidence in SKILL.md: 'Post-Execution Reflection... What failed? — Fix the instruction that caused it.'
  • [DYNAMIC_EXECUTION]: Several phases execute Python scripts via python3 -c using shell variable interpolation to inject file paths and configuration directly into the source code string. This pattern is vulnerable to command injection if a filename or parameter contains shell-breaking characters.
  • Evidence in SKILL.md (Phase 5): uv run --with yake python3 -c "... with open('$INPUT_FILE') as f: ..." where $INPUT_FILE is an unvalidated shell variable.
  • Evidence in references/analysis-tiers.md (Tier 3): uv run --with scikit-learn python3 -c "... chunk_size = $CHUNK_SIZE ...".
  • [COMMAND_EXECUTION]: The skill uses shell variables ($INPUT_FILE, $KEYWORD) inside Bash heredocs and awk commands without sufficient escaping or validation.
  • Evidence in SKILL.md (Phase 6): The awk script uses tolower(kw) where kw is populated from the user-provided $KEYWORD variable.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external .txt recordings (converted from terminal sessions). This content is untrusted and could contain malicious instructions designed to exploit the agent's summary and self-modification logic.
  • Ingestion points: Phase 0 and Phase 1 read external files.
  • Capability inventory: The skill has Bash, rg, and python3 (via uv) capabilities.
  • Sanitization: No sanitization or boundary markers are present to distinguish between recording data and agent instructions.
  • [EXTERNAL_DOWNLOADS]: The skill dynamically installs dependencies from public registries and suggests system package installation.
  • Fetches yake, scikit-learn, and numpy from PyPI via uv run.
  • Recommends installation of ripgrep and fd via Homebrew.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 17, 2026, 06:30 AM