cm
Fail
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONPERSISTENCECOMMAND_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill documentation provides a one-liner installation command that downloads and executes a shell script directly from the author's GitHub repository (Dicklesworthstone/cass_memory_system).
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data sources to extract procedural memory rules, creating a surface where malicious session logs or code comments could influence the agent's future behavior.
- Ingestion points: Processes agent session logs (e.g., .jsonl files in ~/.claude/sessions/) and scans project source code for inline feedback comments (e.g., [cass: helpful/harmful]).
- Boundary markers: The skill uses a multi-stage 'ACE Pipeline' (Generator, Reflector, Validator, Curator) to process data, though specific prompt isolation delimiters are not detailed in the CLI instructions.
- Capability inventory: The tool performs file system operations (read/write), executes shell commands for diagnostics, and communicates with external LLM APIs for pattern extraction.
- Sanitization: Includes a built-in sanitization system that identifies and redacts sensitive patterns such as API keys (Anthropic, OpenAI, AWS, Google), GitHub tokens, and JWTs before processing data.
- [PERSISTENCE]: The skill includes functionality to install persistent integration hooks, including a post-session hook for Claude Code and a Git pre-commit hook (via cm guard --install).
- [COMMAND_EXECUTION]: The tool monitors for and attempts to block destructive system commands (e.g., rm -rf /, DROP DATABASE) through its 'Trauma Guard' safety system, which utilizes a registry of known dangerous patterns.
Recommendations
- HIGH: Downloads and executes remote code from: https://raw.githubusercontent.com/Dicklesworthstone/cass_memory_system/main/install.sh - DO NOT USE without thorough review
Audit Metadata