scratchpad
Pass
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions (SKILL.md) direct the agent to execute dynamically created shell scripts and test suites (e.g., using curl, vitest, or jest) to verify system state and implementation logic.
- [DYNAMIC_EXECUTION]: The provided management script (scripts/scratch.sh) allows for the creation of new executable files (shell scripts, test files) at runtime. This script includes explicit checks to prevent path traversal (blocking '..' and absolute paths) but facilitates the execution of agent-generated logic.
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow where untrusted data from external sources (such as API responses or file contents) could be written to scratch files using the 'write' command. If these files are later executed or read back into the agent's context without sanitization, malicious instructions embedded in the external data could influence the agent. 1. Ingestion points: Content is ingested via the 'write' command in scripts/scratch.sh from stdin, command-line arguments, or local files. 2. Boundary markers: No specific boundary markers or 'ignore' instructions are applied to the content written to scratch files. 3. Capability inventory: The skill enables file writing via scripts/scratch.sh and shell execution as documented in SKILL.md. 4. Sanitization: The management script validates file paths for traversal but does not sanitize or validate the content being written to the files.
Audit Metadata