agent-sdk-builder

Pass

Audited by Gen Agent Trust Hub on Apr 25, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill directly interpolates untrusted user input from {INITIAL_PROMPT} and $ARGUMENTS into its operational instructions, creating a surface for prompt injection.
  • Ingestion points: {INITIAL_PROMPT} variable in SKILL.md and README.md, and $ARGUMENTS in commands/agent-builder.md.
  • Boundary markers: Absent; user input is not delimited from system instructions.
  • Capability inventory: The skill can write files to the workspace (plan/, output/), execute git clone, and launch a web server.
  • Sanitization: None; input is used directly to guide agent behavior and code generation.
  • [EXTERNAL_DOWNLOADS]: The skill fetches resources from official repositories and documentation sites associated with the author.
  • Fetches technical documentation from https://docs.openhands.dev/llms.txt.
  • Clones SDK examples and documentation from the OpenHands GitHub organization.
  • [COMMAND_EXECUTION]: The skill performs shell-based operations to retrieve data and serve generated content.
  • Executes git clone to download external repositories.
  • Generates and executes a local web server to render visual representations of technical plans using HTML, CSS, and Javascript.
  • Generates Python code for custom agents based on user requirements gathered during the interview process.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 25, 2026, 01:31 PM