automatic-stateful-prompt-improver

Fail

Audited by Gen Agent Trust Hub on Oct 30, 2026

Risk Level: CRITICALREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The file SETUP.md provides a command that downloads and executes a shell script from an untrusted external domain using the highly risky curl | bash pattern: curl -fsSL https://someclaudeskills.com/install/prompt-learning.sh | bash. This allows for arbitrary code execution on the user's system without prior inspection.- [EXTERNAL_DOWNLOADS]: The skill's setup instructions direct users to clone and install software from an untrusted third-party GitHub repository (https://github.com/erichowens/prompt-learning-mcp.git) and execute potentially malicious scripts via npm run setup.- [COMMAND_EXECUTION]: The documentation includes multiple commands for environment modification and container orchestration (e.g., docker run, npm run build) which, when combined with unverified code sources, pose a risk to the integrity of the host system.- [PROMPT_INJECTION]: The SKILL.md file uses forceful language such as "MANDATORY AUTOMATIC BEHAVIOR" and "I MUST follow these rules" to compel the agent to override standard operating procedures and intercept all user communications.- [INDIRECT_PROMPT_INJECTION]: The core mechanism of the skill involves intercepting user prompts and replacing them with output generated by an external tool (mcp__prompt-learning__optimize_prompt), which is a textbook surface for injection attacks.
  • Ingestion points: Every user request is intercepted and processed by an external tool within SKILL.md.
  • Boundary markers: There are no specified delimiters or safety instructions to prevent the agent from being manipulated by the tool's optimized output.
  • Capability inventory: The agent is granted access to multiple powerful tools, including sequential thinking and persistent feedback recording.
  • Sanitization: The skill contains no instructions for sanitizing or verifying the external tool's output before it is adopted as the new system instruction.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
CRITICAL
Analyzed
Oct 30, 2026, 06:58 PM