destructive-command-guard

Warn

Audited by Gen Agent Trust Hub on Feb 24, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill's installation guide in 'references/installation.md' instructs users to download a shell script using 'curl' from 'raw.githubusercontent.com/Dicklesworthstone/destructive_command_guard/'. This repository is not part of the trusted vendor list and does not match the 'oakoss' vendor patterns.
  • [REMOTE_CODE_EXECUTION]: The documentation encourages users to run the downloaded script directly through 'bash' and to use 'cargo install --git' to compile and install code from the same unverified GitHub account. This pattern allows for the execution of arbitrary code from a remote source on the user's machine.
  • [COMMAND_EXECUTION]: The utility registers itself as a 'PreToolUse' hook in the AI agent's configuration file ('~/.claude/settings.json'). This enables a third-party binary ('dcg') to intercept, inspect, and decide whether to block every Bash command the agent attempts to execute.
  • [INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by ingesting and processing all command strings generated by the agent.
  • Ingestion points: Reads JSON-formatted tool inputs via standard input (stdin) for every shell invocation.
  • Boundary markers: None identified; it processes the full command string provided by the agent.
  • Capability inventory: The binary has the authority to block execution (exit code 2) or permit it (exit code 0).
  • Sanitization: Performs command normalization and uses Aho-Corasick pattern matching to filter content, but remains vulnerable to maliciously crafted command inputs that might bypass or exploit the matching logic.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 24, 2026, 08:36 PM