skills/alamops/skills/investigate/Gen Agent Trust Hub

investigate

Pass

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external and untrusted sources such as the codebase, git history, and external documentation. This data could potentially contain malicious instructions intended to influence the agent's research findings or the generation of experimental code.
  • Ingestion points: Codebase files, git logs, external library documentation (via web tools), and PRD/spec files (SKILL.md).
  • Capability inventory: Shell command execution (git, gh, cloud CLIs), file creation in the docs/ directory, and the creation/execution of scripts in a scratchpad directory.
  • Boundary markers: The instructions explicitly mandate a "read-only on source" policy and restrict code execution to a "scratchpad" environment.
  • Sanitization: There is no mention of sanitizing or escaping external content before it is processed by sub-agents or included in reports.
  • [DYNAMIC_EXECUTION]: The skill utilizes a "Spikes" mechanism to verify technical assumptions by creating and running small, throwaway scripts at runtime.
  • The instructions direct sub-agents to create a dedicated directory in a scratchpad or temp area, initialize a project (e.g., creating a local package.json), and execute code to answer specific feasibility questions.
  • While the skill emphasizes that these spikes should be isolated, the ability to generate and execute code dynamically constitutes a high-capability attack surface.
  • [COMMAND_EXECUTION]: The skill instructs agents to use various CLI tools to gather repository metadata and environment information.
  • Commands mentioned include git log, git show, gh api, and cloud provider CLIs (e.g., describe or list commands).
  • These tools are used for discovery and research; instructions specify that operations must be read-only and should not modify the environment, data, or deployment state.
  • [REMOTE_CODE_EXECUTION]: The spike mechanism involves initializing local project environments which may download and execute third-party dependencies.
  • The instructions mention creating a local package.json or venv and performing an "install" within the scratchpad to test library behavior.
  • This process may fetch external packages from public registries (npm, PyPI) based on the libraries being investigated, which could lead to the execution of remote code if the libraries or versions identified are malicious.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 3, 2026, 03:37 AM