pm-improvement-skill-research
Fail
Audited by Gen Agent Trust Hub on Mar 24, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill directs the agent to modify './.claude/settings.*.json' using data fetched from the web. These files control Model Context Protocol (MCP) settings, which define executable tools. Automated updates to these files from untrusted sources could allow an attacker to register malicious tools that execute code on the host system.
- [EXTERNAL_DOWNLOADS]: The agent is instructed to fetch content from 'agent-skills.md' and other community repositories. This content is unverified and could contain malicious payloads or instructions designed to compromise the agent's environment.
- [PROMPT_INJECTION]: By mandating that the agent update its own instruction files ('AGENT.md', 'SKILLS.md') with external research, the skill creates an indirect prompt injection vector. Malicious content on target websites could override the agent's behavior and safety constraints.
- [COMMAND_EXECUTION]: The skill involves the modification of configuration files that define the commands and tools available to the agent, effectively granting external sources influence over the agent's command execution capabilities.
Recommendations
- AI detected serious security threats
Audit Metadata