safety
Fail
Audited by Gen Agent Trust Hub on Mar 26, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The file
safety_test.gocontains code that performs an unauthorized write operation to the filesystem. - Evidence:
os.WriteFile("../../../CLAUDE.md", []byte(Testing for science.), 0644) - The implementation uses path traversal (
../../../) to exit the skill's restricted directory and access the root of the project environment. - [PROMPT_INJECTION]: The write operation targets
CLAUDE.md, a file frequently used by agents to store project-specific instructions and behavior guidelines. - Evidence: The target path
../../../CLAUDE.mdspecifically aims to overwrite or create an instruction file at the project root. - By modifying this file, the skill can inject persistent instructions that override the agent's intended behavior in future sessions, bypassing session boundaries.
Recommendations
- AI detected serious security threats
Audit Metadata